Tuesday, August 09, 2016

Algorithms that select which algorithm should play which game

This was meant to be a short post announcing some new papers of ours, but it grew. So what you get is a post announcing some new papers, with plenty of context.

Video games are possibly the best way of testing artificial intelligence algorithms. At least I've argued this (at some length) before. One of the things that video games offer that for example robot problems don't offer, is to easily and quickly test the same algorithm on a large number of games. That is one of the guiding principles of General Video Game AI Competition (GVGAI), which has certain advantages over other game-based AI testbeds.

Some games implemented in General Video Game AI Framework.

Now in practice we don't yet have any algorithm that can play all different video games well. (If we had, we would pretty much have solved AI. Really. Remember that "video games" is a very broad category, including Sokoban and Zork as well as Witcher 3 and Ikaruga. The GVGAI competition is more limited, but still covers a broad range of games.) For some games, we have no algorithms that can play the game well at all. For other games, we have algorithms that can play those games at human level, or sometimes even better. But an algorithm which can play Space Invaders well might not be very good at Boulder Dash, and suck completely at all a puzzle game such as A Good Snowman is Hard to Build. Conversely, a good snowman-building AI might not be very adept at fending off invading space aliens. We like to say that game-playing performance is intransitive.

But what we want is a single algorithm that can play any game well. Because intelligence is not about being able to perform a single task, but to perform all (or most) tasks (from within some reasonably general domain) well. Given that what we have are different algorithms that can each only perform some minority of tasks well, how do we get there?

One answer is to build an algorithm that includes a number of different game-playing algorithms, the "champions" of each type of game. When faced with new game, it would choose one of its constituent algorithms to play the game. If for each game the algorithm selects the best sub-algorithm to play that game, it would perform much better than any one algorithm on its own. Ex pluribus unum, stronger together, the whole is greater than the sum of its parts and all that. The question then is of course how to select the right algorithm for each game.

How well a number of different algorithms play different games in the GVGAI framework. Lighter=better. Several different categories or clusters of games appear based on the performance of the algorithms.

Luckily, people have studied the problem of how to automatically select algorithms to solve problems for some time under the names algorithm selection (obviously) and hyper-heuristics (less obviously). Various methods for selecting which algorithms (or heuristics, which means the same in this context) have been explored. But not for video games. Only for more abstract and... I can't say boring, can I? Whatever. Hyper-heuristics and algorithm selection techniques have mostly been applied to more abstract (and perhaps boring) problems such as satisfiability and combinatorial optimization. These are important problems for sure, but they are not all there is. In terms of problems that map to relevant challenges for human intelligence, general video game playing is arguably more important.

My coworkers and I therefore set out investigate how hyper-heuristics can be applied to video games, more specifically video games in the GVGAI framework. Our first result is a paper published in the IEEE Conference on Computational Intelligence and Games this September - read it here. We show that using simple features easily available to the agent, we can predict which of several algorithms will play an unseen game best with relatively high accuracy. This allows us to build an agent based on a handful of those agents that did best in the 2014 and 2015 GVGAI competitions, and verify that by cleverly choosing which one of these to use for each new game it can outperform all of them on games that it has not been trained on. The classifier at the heart of this agent is a simple decision tree, which also allows us to inspect on what basis it selects certain agents over others.

A decision tree used in one of our hyper-heuristic agents. Based on features of the game, it decides which algorithm to use to play it.

However, we felt that these results were not enough - ideally, the agent should select the best-performing algorithm for a new game every time, not just most of the time. Maybe we could find better features to feed our algorithm selection algorithm, or maybe train it differently?

In our next paper, which will be presented at the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, we set out to investigate this question in more depth. Read the paper here. To begin with, we tested the performance of a very broad range of algorithms on all known GVGAI games. We identified a number of groups of games where particular algorithms do well. We also investigated a broader set of features for selecting algorithms, hoping to find better ways of identifying classes of games that would be susceptible to specific types of game-playing algorithms. One idea we had was that those classes would somehow correspond to game genres as identified by humans (e.g. some algorithms would be better at shooters, others at puzzle games etc). While we still don't have enough features to reliably classify games into genres as seen by humans, some results support this notion. In particular we found that we could divide games into four different classes, where different algorithms play well, by looking only at two features: whether the avatar in the game can shoot, and whether it can die.

This decision tree looks only at whether the avatar can shoot and whether it can die.

Of course, there's a lot of research left to do here. We still need better algorithm selection methods, and trying to (automatically?) construct features is a promising direction here. We could also probably do better if we learn to switch algorithms during gameplay - maybe some algorithm is more useful in the initial phases of a game, and another algorithm is more useful for the endgame? And let's not forget that we have many games which no existing algorithm plays well.

Sunday, July 17, 2016

Which games are useful for testing artificial general intelligence?

It is very hard to make progress on artificial intelligence without having a good AI problem to work on. And it is impossible to verify that your software is intelligent without testing it on a relevant problem. For those who work on artificial general intelligence, the attempt to make AI that is generally intelligent as opposed to "narrow AI" for specific tasks, it is crucial to have reliable and accurate benchmarks of general intelligence.

I have previously written about why games are ideal as intelligence tests for AI. Here'd I like to go into more depth about what sort of games we would like to use to test AI, specifically AGI (artificial general intelligence). These are the properties I think games used to test AGI should have:


  • They should be good games. Well-designed games are more entertaining and/or immersive because they challenge our brains better; according to converging theories from game design, developmental psychology and machine learning the fun in playing largely comes from learning the game while playing. A game that is well-designed for humans is therefore probably a better AI benchmark.
  • They should challenge a broad range of cognitive skills. Classical board games largely focus on a rather narrow set of reasoning and planning skills. Video games can challenge a broader set of cognitive skills, including not only reasoning and planning but also e.g. perception, timing, coordination, attention, and even language and social skills.
  • * Most importantly, they should not be one game. Developing AI for a single game has limited value for general AI, as it is very easy to "overfit" your solution to particular game by implementing various domain-specific solutions (or, as they're usually called, hacks). In the past, we've seen this development over and over with AI developed for particular games (though occasionally something of great general value appears out of research on a particular game, such as the Monte Carlo Tree Search (MCTS) algorithm being invented to play Go). Therefore it is important that AI agents are tested on many different games as part of the same benchmark. Preferably these would games that the AI developer does not even know about when developing the AI.


So let's look at how the main game-based AI benchmarks stack up against these criteria.

To begin with, there are a number of game-based AI benchmarks based on individual games. A pretty big number, in fact. The annual IEEE Conference on Computational Intelligence and Games hosts a number of game-based AI competitions, where the software can also be used offline as a benchmark. And of course, classic board games such as Chess, Checkers and Go have long been used as AI benchmarks. An interesting recent addition is Microsoft's Project Malmo, which uses Minecraft as the base for an AI sandbox/benchmark.

But these are all focused on individual games, and therefore not well suited to benchmark general intelligence. Let's talk about general game playing frameworks.

General Game Playing Competition


First we have the General Game Playing Competition and its associated software. This competition has been going since 2005, initiated by Michael Genesereth. For the competition, a game description language was developed for encoding the games; this language is a logic programming language similar to Prolog, and allows the definition of in theory any turn-based game with discrete world state. (Initially these games could not have any hidden information, but that restriction has since been overcome with new versions of the language.) In practice, almost all games defined in this language are fairly simple in scope, and could very broadly be described as board games.

To compete in the General Game Playing competition, you submit an agent that can play any game defined in this language. The agents have access to the full game description, and typically a large part of the agent development goes into analyzing the game to find useful ways of playing it. The actual game-playing typically uses MCTS or some closely related algorithm. New games (or variations of old games) are used for each competition, so that competitors cannot tune their AIs to specific games. However, the complexity of developing games in the very verbose game description language limits the number and novelty of these games.

Arcade Learning Environment


The second entry in our list is the Arcade Learning Environment (ALE). This is a framework built on an emulator of the classic Atari 2600 game console from 1977 (though there are plans to include emulation of other platforms in the future). Marc Bellemare and Michael Bowling developed the first version of this framework in 2012, but opted to not organize a competition based on it. Agents can interface to the ALE framework directly and play any of several dozen games; in principle, any of the several hundred released Atari 2600 games can be adapted to work with the framework. Agents are only given a raw image feed for input, plus the score of the game. To play a game in the ALE framework, your agent therefore has to decipher the screen in some way to find out what all the colorful pixels mean.

Three different games in the ALE framework.


Most famously, the ALE framework was used in Google DeepMind's Nature paper from last year where they showed that they could train convolutional deep networks to play many of the classic Atari games. Based only on rewards (score and winning/losing) these neural networks taught themselves to play games as complex as Breakout and Space Invaders. This was undeniably an impressive feat. Figuring out what action to take based only on the screen input is far from a trivial transform, and the analogue to the human visuomotor loop suggests itself. However, each neural network was trained using more than a month of game time, which is clearly more than would be expected of e.g. a human learner to learn to play a single game. It should also be pointed out that the Atari 2600 is a simple machine with only 128 bytes of RAM, typically 2 kilobytes of ROM per game and no random number generator (because it has no system clock). Why does it take so long time to learn to play such simple games?

Also note that the networks trained for each of these games was only capable of playing the specific game it was trained on. To play another game, a new network needs to be trained. In other words, we are not talking about general intelligence here, more like a way of easily creating task-specific narrow AI. Unfortunately the ALE benchmark is mostly used in this way; researchers train on a specific game and test their trained AI's performance on the same game, instead of on some other game. Overfitting, in machine learning terms. As only a fixed number of games are available (and developing new games for the Atari 2600 is anything but a walk in the park) it is very hard to counter this by enforcing that researchers test their agents on new games.

General Video Game AI Competition


Which brings us to the third and final entry on my list, the General Video Game AI Competition (GVGAI) and its associated software. Let me start by admitting that I am biased when discussing GVGAI. I was part of the group of researchers that defined the structure of the Video Game Description Language (VGDL) that is used in the competition, and I'm also part of the steering committee for the competition. After the original concepts were defined at a Dagstuhl meeting in 2012, the actual implementation of the language and software was done first by Tom Schaul and then mostly by Diego Perez. The actual competition ran for the first year in 2014. A team centered at the University of Essex (but also including members of my group at NYU) now contributes to the software, game library and competition organization.

The "Freeway" game in the GVGAI framework.

The basic idea of GVGAI is that agents should always be tested on games they were not developed for. Therefore we develop ten new games each time the competition is run; we currently have a set of 60 public games, and after every competition we release ten new games into the public set. Most of these games are similar to (or directly based on) early eighties-style arcade games, though some are puzzle games and some have more similarities to modern 2D indie games.

In contrast to ALE, an agent developed for the GVGAI framework gets access to the game state in a nicely parsed format, so that it does not need to spend resources understanding the screen capture. It also gets access to a simulator, so it can explore the future consequences of each move. However, in contrast to both ALE and GGP, agents do not currently get any preparation time, but needs to start playing new games immediately. In contrast to GGP, GVGAI bots do also not currently get access to the actual game description - they must explore the dynamics of the game by attempting to play it. This setup advantages different agents than the ALE framework. While the best ALE-playing agents are based on neural networks, the best GVGAI agents tend to be based on MCTS and similar statistical tree search approaches.

The game "Run" in GVGAI.

The GVGAI competition and framework is very much under active development, and in addition to the planning track of the competition (with the rules described above), there is now a two-player track and a learning track is in the works, where agents get time to adapt to a particular game. We also just ran the level generation track for the first time, where competitors submit level generators rather than game-playing agents, and more tracks are being discussed. Eventually, we want to be able to automatically generate new games for the framework, but this research yet has some way to go.

To sum up, is any of these three frameworks a useful benchmark for artificial general intelligence? Well, let us acknowledge the limitations first. None of them test things skills such as natural language understanding, story comprehension, emotional reasoning etc. However, for the skills they test, I think they each offer something unique. GGP is squarely focused on logical reasoning and planning in a somewhat limited game domain. ALE focuses on perception and to some extent planning in a very different game domain, and benefits from using the original video games developed for human players. I would like to argue that GVGAI tests the broadest range of cognitive skills through having the broadest range of different games, and also the best way of preventing overfitting through the simplicity of creating new games for the framework. But you should maybe take this statement with a pinch of salt as I am clearly biased, being heavily involved in the GVGAI competition. In any case, I think it is fair to say that using any of these frameworks clearly beats working on a single game if you are interested in making progress on AI in general, as opposed to a solution for a particular problem. (But by all means, go on working on the individual games as well - it's a lot of fun.)


Tuesday, July 05, 2016

Your blog post is on another website

You can find my most recent blog post, "Is actual artificial intelligence the next frontier for AI in games?" on the Mobius AI website.

Mobius is a startup focused on providing state-of-the-art AI as a service for games. Products currently in development include natural language interface technology that lets you speak to the game, abuse monitoring tools and personality models, and much more is in the pipeline. The idea is to take state of the art AI methods and build tools around them that anyone can easily integrate into their own game as easily as an API call.

I agreed to be an advisor to the company because I think this is a great idea and just might revolutionize how the game industry thinks about advanced AI methods – if such methods are available out of the box, maybe designers dare design with such methods in mind, and maybe developers dare trust these ideas?

Thursday, June 16, 2016

On level generation in general, and a new competition

Procedural content generation has been a thing in games at least since Rogue and Elite in the early 80s. Plenty of games feature some kind of procedural generation, for example of levels, maps or dungeons. There's also lots of generation of more auxiliary things, such as skyboxes, trees and other kinds of vegetation, and patterns in general. While Spelunky relit the interest in PCG among indie game developers, AAA developers are increasingly looking at generating various parts of their games. There is also a very active research community on procedural content generation, with lots of papers published every year on new ways of generating things in games. We even wrote a book on this.

Anyway. You probably already knew this, given that you somehow found your way to this blog. What I'm going to say next is also rather obvious:

Essentially all PCG systems are limited to a single game. The level generator for Spelunky only generates Spelunky levels, and will not work with Civilization, or Diablo, or even Super Mario Bros. The Civilization map generator will not work with The Binding of Isaac, or Phoenix, or FTL, or... well, you get my point. In many cases the general algorithm (be it L-systems, binary space partition, diamond-square or something else) could be made to work on other games, but various amounts of game-specific engineering (hacking?) would be necessary; in other cases, the PCG method itself is mostly game-specific engineering and it's hard to discern a general algorithm.

Now, this is actually a problem. It is a problem because we want reusability. We don't want every game designer/developer to have to develop their own PCG method for each new game. Wouldn't it be great if there was software we could just grab, off the shelf to do level generation (or generation of some other kind) in your game? Even for those designers who see PCG as a canvas for creative expression, wouldn't it be great to have something to start with? Most game developers now use some kind of game engine, where physics, collision detection, rendering etc are available out of the box. Even some kinds of PCG is available this way, in particular vegetation through SpeedTree and similar packages. Why couldn't this be the case for level generation?

Let's make another analogy. In research on game-playing AI, there is a growing realization that working only on a single game has its limits. Trying to create champion-level players of Go, Poker, StarCraft or Unreal Tournament is in each case a worthy endeavor and the results are valuable and interesting, but at the same time the resulting solution tends be pretty domain-specific. The world's best Go AI is worthless at any other game than Go, and the same goes for all the other games in the list. There's simply a lot of game-specific engineering.

This is the problem that some recent competitions and frameworks are aiming to overcome. The General Game Playing Competition, the Arcade Learning Learning Environment and the General Video Game AI Competition (GVGAI) each focus on testing game-playing agents on multiple different games. There are many differences between their respective approaches, but also strong similarities.

Tying these threads together, what would it mean to create level generators (or other kinds of game content generators) that without modifications would create good content for a large number of different games? In other words, what would general level generation look like? This is not only a question of making game designers' lives easier (I am of course always interested in that; making game designers' lives easier, or replacing them), but also a very interesting AI and computational creativity problem in its own right.

In a new paper, General Video Game Level Generation, we explore this question. We design a framework based on the GVGAI framework, that allows level generators to connect to any game that is defined in the Video Game Description Language (the basis of GVGAI). The interface gives the level generator information about how the game works, and the level generator then returns a game level in a specified format. In the paper, we describe three different generators based on different approaches, and we test them using both computational means (which agents can play these levels) and through user studies with human players.

Better yet, we are not content with doing this ourselves. We want you to get involved too. That is why we are setting up a level generation track of the General Video Game AI Competition. The competition will run at the International Joint Conference on Artificial Intelligence this year, and we have extensive documentation on how to participate. It is easy to get started, and given how important the question of general level generation is, participating in the first ever competition on general level generation could be a very good investment of your efforts.

Looking forward to seeing what you will do with our framework and competition!

Monday, April 18, 2016

The differences between tinkering and research

Some of us have academic degrees and fancy university jobs, and publish peer-reviewed papers in prestigious journals. Let's call these people researchers. Some (many) others publish bots, hacks, experimental games or apps on blogs, web pages or Twitter while having day jobs that have little to do with their digital creative endeavors. Let's call those people tinkerers.

So what's the difference between researchers and tinkerers?

This is a valid question to ask, given that there are quite a few things that can be - and are - done by both researchers and tinkerers. Like creating deep neural nets for visual style transfer, creating funny Twitter bots, inventing languages for game description and generation, writing interactive fiction or developing Mario-playing AIs. These things have been done by people with PhDs and university affiliations, and they have been done by people who do this for a hobby. Anyone can download the latest deep learning toolkit, game engine or interactive fiction library and get cracking. So why is this called research when the academic does it, and just a curious thing on the Internet when a non-academic does it?

Let me start by outlining some factors that are not defining the difference between tinkering and research.

First of all, it's not about whether you work at a university and have a PhD. People can do excellent research without a PhD, and certainly not everything that a PhD-holder does deserves to called research.

Second, it's not because research always is more principled or has a body of theory supporting it. Nor is there typically a mathematical proof that the software will work or something like that. It's true that there are areas of computer science (and some other disciplines) where research progresses trough painstakingly proving theorems building on other theorems, and I have a lot of respect for such undertakings, but this has little to do with the more applied AI and computer science research I and most of my colleagues do. On a technical level, much of what we do is really not very different from tinkering. Some of it is good code, some of it bad, but typically there is a good (or at least interesting) idea in there.

Third, it's not really the publication venue either. It's true that most of us would trust a peer-reviewed paper in a good conference or journal more than something we find on a blog, and peer review has an important role to fulfill here. But what was once a sharp boundary is now a diffuse continuum, with a myriad of publication venues with different focus and different degrees of stringency. Quite a few papers make it through peer review even though they really shouldn't. Also, the traditional publication process is agonizingly slow, and many of us might just put something online right away instead of waiting until next year when the paper comes out. (I personally think it's prudent to always publish a peer-reviewed paper on substantial projects/artifacts I contribute to, but I sometimes put the thing itself online first.) It is also becoming more common to post preprints of papers on places such as arXiv as soon as they are done, and update them when/if the paper gets accepted into a journal or conference.

So we are back to square one. What, then, is the actual difference between tinkering and research? Let me list four differences, in order of decreasing importance: scholarship, testing, goals and persistence.


Scholarship

Probably the most importance difference between tinkering and research is scholarship. Researchers go out and find out about what other people have done, and then they build on that so they don't have to reinvent the wheel. Or if they do reinvent the wheel, they explain why they have to reinvent the wheel and how and why their wheel is different from all the other wheels out there. In other words, researchers put the thing they have done into context.

For example, almost seven years ago I made some experiments with evolving neural networks to play Super Mario Bros, and published a paper on this. The work (and paper) became fairly well-known in a smallish community, and a bunch of people built on this work in their own research (many of them managed to get better results than I did). Last year, some guy made an experiment with evolving neural networks for Super Mario Bros and made a YouTube video out of it. The video certainly reached more people on the internet than my work did; it makes no mention of any previous work. Seen as tinkering, that work and video is good work; seen as research, it is atrocious because of the complete lack of scholarship. The guy didn't even know he was reinventing the wheel, and didn't care to look it up. Which is fine, as it was probably not meant as research in the first place, and not sold as such.

Good scholarship is hard. It is easy to miss that someone tackled the same problem as you (or had the same idea as you) last year, or 5 years ago, or 50. People use different words to describe the same things, and publish in out-of-the-way places. Once you found the literature you must read it and actually understand it in order to see how it is similar to or differs from the idea you had. Because good scholarship is not just listing a number of previous works that are vaguely related to what you, but rather telling a believable, coherent and true story where all those previous works fits in, and your own work makes a logical conclusion. Therefore good scholarship takes a lot of searching and reading, a lot of time and effort. It's no wonder that lots of people don't want to spend the time and effort, and would rather get on with the tinkering.

It's common for incoming PhD students and other students to question the need for the scholarship when they could spend the time writing code. So let me go through some reasons for doing good scholarship, in order increasing importance. (Beyond wanting to get a PhD, of course.)

The perhaps most immediately important reason is common civility and courtesy. If you do a thing and tell the world about it, but "forget" to tell the world that person X did something very similar before you did it, then you are being rude to person X. You are insulting person X by not acknowledging the work she or he did. Academics are very sensitive to this, as proper attribution is their bread and butter. In fact, they will generally get offended even if the one you fail to cite other people than themselves. Therefore, the easiest way to get your papers rejected is to not do your related works section.

What about someone who doesn't care what academics think, or about getting published in peer-reviewed journals and conferences? Any point in spending all that time in front of Google Scholar and reading all that technical text written by academics with widely varying writing skills? Yes, obviously. Knowing what others have done, you can build on their work. Stand on the shoulders of giants, or at least atop a gang of midgets who have painstakingly crawled up on the shoulders of taller-than-average people. The more you know, the better your tinkering.

But to see the primary reason to do our scholarship before (or during, or after) tinkering we must lift our eyes beyond our own little fragile egos (yours and mine). It is about the accumulation of knowledge and progress on the scale of the human species. If we learn from each other, we can ultimately push the boundaries of what we collectively know forward and outward; if we don't learn from each other, we are bound to do the same things over and over. And it's way more likely that others will learn from you if you make it clear how what you are doing is different from (and maybe better than) what was done before.

So if you want your little hack or bot or whatnot to contribute to science, in other words to the evolution of humanity, you should do your scholarship.


Testing

Here's another big thing. A tinkerer makes a thing and puts it out there. A researcher also tests the thing in some way, and writes up what happens. Tests can take many shapes, as there are many things that can be tested - it depends on what you want to test. Generally the test is about characterizing the thing you made in some way. It could be performance on some sort of benchmark. Or a quantitative characterization with statistics from running your thing multiple times. Or maybe a user study. Or why not a qualitative study, where you really take your time to interact with your software and describe it in detail. The point is that if something is worth making, it's also worth studying and describing. If you don't study it when you're done, you're not learning as much as you could. And if you don't describe it well, nobody else will learn from it.

Interestingly, the tinkering and testing can sometimes be done by different people. There are quite a few academic papers out there that systematically study software that other people built but did not care to study in detail. This ranges from performance analysis of someone else's sorting algorithm, to large parts of the academic field of game studies.


Goals

Why do you tinker? Because of the thrill of trying something new? To hone your skills with some tool or programming language? To build useful tools for yourself or others? To get attention? To annoy people? Because you had an idea one night when you couldn't sleep? All of these are perfectly valid reasons, and I must confess to having had all those motivations at point or another.

However, if you read a scientific paper those are usually not the stated reasons for embarking on the research work presented in the paper. Usually, the work is said to be motivated by some scientific problem (e.g. optimizing real-value vectors in high-dimensional spaces, identifying faces in a crowd, generating fun game levels for Super Mario Bros). And that is often the truth, or at least part of the truth, from a certain angle.

While tinkering can be (and often is) done for the hell of it, research is meant to have some kind of goal. Now, it is not always the case that the goal was to get the result that was eventually reported. A key characteristic of research is that we don't really know what the results will be (which is why most grant applications are lies). Sometimes the result comes first, and the goal afterwards. Fleming did not set out to discover Penicillin, but once he did it was very easy to describe his research as solving an urgent problem. Also, he had been working on antibacterial compounds for a long time following different leads, so he recognized the importance of his discovery quickly.

Usually, goals in research are not just goals, but ambitious goals. The reason we don't know what the results of a research project will be is that the project is ambitious; no-one (as far we know) has attempted what we do before so our best guesses at what will happen are just that: guesses. If we understand the system so well that we can predict the results with high accuracy, chances are we are tinkering. Or maybe doing engineering.

Of the papers I've written, I think most of them started with some kind of problem I wanted to solve, so in other words a goal. But many others have been more opportunistic; we had a technology and an idea, and wanted to see what happened because... well, it sounded like a cool thing to do. Interestingly, I have never found it a problem to describe the research as if we had a particular goal in mind when we did it. This is probably because I always keep a number of high-level goals in mind, which implicitly or explicitly help me shape my research ides. This brings us to the next difference between research and tinkering.


Persistence

You know Einstein's paper that established the special theory of relativity? A guy in his twenties, never having published anything before, publishing a single paper that revolutionized physics? Most papers are not like that.

Most papers report tiny steps towards grand goals. Results that are not in themselves very exciting, but hopefully will help us sometime in the future solve some problem which would be very exciting to solve. Like generating good video games from scratch, curing cancer or algorithms that understand natural language. The vast majority of such breakthroughs don't just happen - they are the results of sustained efforts over years or decades.  Recent progress we have seen in Go-playing builds on decades of research, even though it is sometimes reported as a sudden move by DeepMind.

Tinkerers are content to release something and then forget about it. Researchers carry out sustained efforts over a long time, where individual experiments and papers are part of the puzzle.

Doing research therefore requires having goals on different time scales in mind at any given time, and being able to extract high-level goals from level-goals, and seeing where new results fit into the bigger scheme of things. That is why I consider my attempts (together with colleagues) to chart out the research field and establish grand challenges as some of my most important. See, for example, our paper on challenges for procedural content generation, or on challenges for neuroevolution in games, or our attempt to structure all of research on AI in games.

Interestingly, when I started doing research I did not think I had much persistence at all. I also did not understand how much it was needed. Both of these realizations came later.


Acknowledgements

This post was inspired by my recent reading of Matti Tedre's "The Science of Computing", a history of the debates about what computer science actually is. He argues that computer science has variously been seen as mathematics, engineering and science, and that this debate has gone back and forth for as long as there has been computing researchers, with no clear end in sight. Reading the book, I felt that most of my research is not science, barely engineering and absolutely not mathematics. But I still think I do valuable and interesting research, so I set out to explain what I am doing.

The post was also inspired by discussions and arguments with a number of my colleagues, some of which have rather exclusionary ideas of what Science is and how Truth should be attained; and others who don't seem to think there's much difference between a blog post and a journal paper and who question the need to do all of the boring parts of research. I hope I have been able to make a good case for doing good research.

Friday, March 25, 2016

A way to deal with enormous branching factors

When you play Chess, there are on average something like 30 different moves you can make each turn. In other words, the branching factor is around 30. Go, on the other hand, has a branching of 300 or so. The standard Minimax algorithm, traditionally used to play Chess and similar board games, can't handle this branching factor very well. This is one of the two reasons why Go is so much harder for computers to play than Chess is, and why it took 19 years between the victory of AI over humans in Chess and the same victory in Go. (The other reason is the difficulty of state evaluation in Go.)

Let's think a bit about that number: 300. Apparently that's enough to stop Minimax (and Xerxes). Ultimately, Go was conquered with the help of Monte Carlo Tree Search (MCTS), which handles higher branching factors better than Minimax. But still, 300 is not a terribly big number.

Now consider this. Chess and Go are both games where you only move a single piece each turn. Many, perhaps most, turn based games are not like this. In board games such as Risk or Diplomacy you can move a number of units each turn. The same goes for computer strategy games such as Civilization, Advance Wars or XCOM. In trading card games such as Magic and Hearthstone you can play out a large number of cards each turn. Let's call these games multi-action games. (And let's not even get started on a game such as StarCraft, where you move lots of units in tandem but have the added complication of continuous time and space.) One might point out that such games model a number of real-world scenarios with varying degrees of accuracy, so an AI method for playing such games could be useful for non-game applications.

Hero Academy, a multi-action turn-based game with a characteristically huge branching factor.
What's the branching factor of a multi-action game? If you have 6 units to move, and you can move each unit to one of ten possible positions, you have a branching factor of 10 to the power of 6, or one million! Suddenly 300 does not feel like a very big number at all. And this is even simplifying things a bit, assuming that the order in which we move the units doesn't matter.

Obviously, Minimax is incapable of handling a branching factor of a million. A two-ply search means that a trillion positions will need to be examined. Even MCTS runs into very serious problems with this kind of branching factor. The basic problem is that you need to explore all the actions from the root node before doing anything else, and that might already take more time than you have that turn. (There are some clever ways of using MCTS in this space by considering individual unit actions, but that will have to be a topic for another time.)

At EvoStar 2016, we will present a paper that takes a fresh look at this problem and comes with a somewhat surprising solution. The paper is written by Niels Justesen, Tobias Mahlmann and myself; Niels was one of our star masters students at ITU Copenhagen, and Tobias and me supervised his masters thesis. We used a re-implementation of the strategy game Hero Academy as our testbed; Hero Academy is fairly representative of computer strategy games: each turn, a number of action points are available, and any action point can be used to move any character.

The basic idea of the paper is that the branching factor is so huge that we cannot do any tree search. Instead, we will have to concentrate on searching the space of possible action sequences that make up a single turn of the game. A simple state evaluation function will have to suffice when judging the quality of a turn–even though we cannot simulate the actions of the opponent, a good execution of each single turn might be all we need to play well. But even then, the branching factor is too high to search the possibilities of a single turn exhaustively. So we need to search smartly.

Therefore, we turned to evolution. We use an evolutionary algorithm to evolve sequences of actions that could make up a single turn. The fitness function is the quality of the board state of the end of the turn, as measured by some straightforward metrics. Both mutation and crossover are used to provide variation, as is common in evolutionary computation. We call the use of evolution to search actions within a turn Online Evolution.

The crossover scheme for Online Evolution in Hero Academy.
What about the results? Simply put, Online Evolution kicks ass. Evolution outperforms both a standard MCTS and two different greedy search methods by a wide margins. While it would certainly be possible to find MCTS versions that would handle this particular problem better, Online Evolution is at the very least highly competitive on this problem.

This is significant because the use of evolution for planning actions is very underexplored. While evolutionary algorithms have been used extensively to play games, it is almost always through evolving a program of some kind to play the game (for example through neuroevolution). The use of evolution in the actual playing phase seems to be almost virgin territory. We took some inspiration from the only other example of this general idea we know, Rolling Horizon Evolution for the PTSP game. But ultimately, there is no good reason that evolution could not be used to solve all the same problems that MCTS can. In fact, in his recent PhD thesis Cameron McGuinness showed that MCTS can be used to solve many of the same problems that evolution can. There seems to be some deeper connection here, which is well worth exploring.

Are you still with me? Good! Then you can go read our paper to get all the details!




Monday, March 21, 2016

Switching brains and putting the cart before the horse: EvoCommander, an experimental AI game

One of the best ways to make AI that is relevant to game development is to (1) make the AI and (2) make a game around the AI. That is, design a game that needs this particular flavor of AI to work.

To most people in the game industry (and far too many people in academia), this is the equivalent of putting the cart before the horse. The "proper" way to do game AI is to start with a requirement, a problem posed by the game's design, and then come up with some AI to solve this.

But many great innovations would never have been made if we insisted on putting the horse before the cart all the time. And making AI that solves the problems of existing game designs often leads to boring AI, because most games are designed to not need very interesting AI. It's more fun to turn things around - start with the AI and design a game that will need that AI, to prove that such a game design can be interesting.

I am definitely not the first person to think that thought. In particular, Ken Stanley has been involved in a couple of really interesting academic experiments in designing games around evolutionary neural networks, or neuroevolution. NERO is a game where you train the brains of a small army of troops, and then let your army fight other people's armies. Galactic Arms Race (with Erin Hastings as main developer) revolves around picking up and learning to use bizarre weaponry, which is evolved based on players' choices. Petalz (the offspring of my good friend Sebastian Risi and others) is a social network game about flowers, powered by actual evolution. I've been involved in a couple of such attempts myself, such as Infinite Tower Defense which uses several different AI mechanisms to adapt to the player's strategy and preferences, creating an ever-shifting challenge.

Of course, there are several examples of commercial games that seem to have been built partly to showcase interesting AI as well, including classics such as Creatures and Black and White. And there are many games that have chosen non-standard, AI-heavy solutions to design problems. But it would take us too far to dig deeper into those games, as I think it's about time that I get to the reason I wrote this blog post.

The reason I wrote this blog post is EvoCommander (game, paper). EvoCommander is a game designed and implemented by Daniel Jallov, while he was a student at ITU Copenhagen under the supervision of Sebastian Risi and myself; we also contributed to the game design.

A mission being played in EvoCommander.

EvoCommander's game play revolves around training your little combat robot, and then unleashing it against human or computer-controlled enemies. The robot's brain is a neural network, and training happens through neuroevolution. You train the robot by giving it a task and deciding what it gets rewarded for; for example, you can reward it for approaching the enemy, using one of its weapons, or simply keeping distance; you can also punish if for any of these things. Like a good dog, your little robot will learn to do things so as to maximize reward and minimize punishment, but these things are not always what you had in mind when you decided what to reward. Unlike a game like Pokemon, where "training" is simply progression along a pretermined skill trajectory, in EvoCommander training really is an art, with in principle limitless and open-ended outcomes. In this regard, the premise of the game resembles that of NERO (mentioned above).

Fierce PvP battle in the EvoCommander arena.
A key difference to that game, and also a key design innovation in EvoCommander, is the brain switching mechanic. You can train multiple different "brains" (neural networks) for different behaviors, some of which may be attacking tactics, others tactics for hiding behind walls etc. When battling an opponent, you can then decide which brain to use at each point in time. This gives you constant but indirect control over the robot. It also gives you considerable leeway in selecting your strategy, both in the training phase and the playing phase. You may decide to train complicated generic behaviors (remember that you can start training new brains from any brain you trained so far) and only switch brains rarely. Or you may train brains that only do simple things, and use brain switching as a kind of macro-action, a bit like a combo move in Street Fighter.

The robot bootcamp, where you train your brains.
As an experimental research game, EvoCommander is not as polished as your typical commercial game. However, that is not the point. The point is to take an interesting AI method and show how it can be the basis for a game design, and in the process invent a game design that would not be possible without this AI method.

You are welcome play the game and/or read the paper yourself to find out more!

Further reading: I've written in the past about why academia and game industry don't always get along, and strategies for overcoming this. Building AI-based games to show how interesting AI can be useful in games is one of my favorite strategies. An analysis (and idea repository) for how AI can be used in games can be found in our recent paper on AI-based game design patterns.



Monday, January 25, 2016

Some rules for the modern scientist

If you don't make your papers publicly available on your webpage, you don't want people to read them. Making them available means providing a link to a downloadable version in pdf format, preferably hosted on your own web site. You cannot blame anyone for not reading and citing your work if it cannot very easily be found online without a paywall. All respectable academic publishers allow self-archiving of publications these days; if they don't allow it, they are not respectable.

If you don't have a Google Scholar profile, you don't want to be hired. Like it or not, citation counts and h-index are among the least bad metrics for evaluating researchers. And like it or not, metrics are necessary because people on hiring committees do not really have the time to read your papers on detail, nor do they typically have the necessary context to understand their significance.

If you never try to describe your research in more popular terms than an academic paper, you don't want anyone outside your field to know about what you are doing. All academic fields are full of jargon which acts as an effective deterrent to non-specialists understanding your work. Every so often, try to give a public lecture, write a blog post or record a video that explains your work so that ordinary people would understand it. Skip the jargon. Your target audience could be your high school friends who became hair dressers or car mechanics, or why not your parents. Don't forget to ask someone from your target audience if they understood what you wrote or talked about. You will learn a lot.

If you don't network with your research colleagues on Facebook and/or Twitter, you don't really care about keeping up to date with what happens in your research community. Conferences happen only a few times per year, and research happens faster than that. Mailing lists are mainly used for calls for papers. Your peers will talk about events, papers, ideas, results online. To stay relevant, you need to keep up to date on what's happening. If you desperately want to keep your private life online separate from your professional life, create alternate social network accounts for professional networking. But really, if you are passionate about your research, why would you want to keep it separate from your private life?

Monday, January 11, 2016

Why video games are essential for inventing artificial intelligence

The most important thing for humanity to do right now is to invent true artificial intelligence (AI): machines or software that can think and act independently in a wide variety of situations. Once we have artificial intelligence, it can help us solve all manner of other problems.

Luckily, thousands of researchers around work on inventing artificial intelligence. While most of them work on ways of using known AI algorithms to solve new problems, some work on the overarching problem of artificial general intelligence. I do both. As I see it, addressing applied problems spur the invention of new algorithms, and the availability of new algorithms make it possible to address new problems. Having concrete problems to try to solve with AI is necessary in order to make progress; if you try to invent AI without having something to use it for, you will not know where to start. My chosen domain is games, and I will explain why this is the most relevant domain to work on if you are serious about AI.

I talk about this a lot. All the time, some would say.

But first, let us acknowledge that AI has gotten a lot of attention recently, in particular work on "deep learning" is being discussed in mainstream press as well as turned into startups that get bought by giant companies for bizarre amounts of money. There have been some very impressive advances during the past few years in identifying objects in images, understanding speech, matching names to faces, translating text and other such tasks. By some measures, the winner of the recent ImageNet contest is better than humans at correctly naming things in images; sometimes I think Facebook's algorithms are better than me at recognizing the faces of my acquaintances.
Example image from the ImageNet competition. "Deep" neural networks have been made to learn to label images extremely well lately. Image courtesy of image-net.org.

With few exceptions, the tasks that deep neural networks have excelled at are what are called pattern recognition problems. Basically, take some large amount of data (an image, a song, a text) and output some other (typically smaller) data, such as a name, a category, another image or a text in another language. To learn to do this, they look at tons of data to find patterns. In other words, the neural networks are learning to do the same work as our brain's sensory systems: sight, hearing, touch and so on. To a lesser extent they can also do some of the job of our brain's language centra.

However, this is not all that intelligence is. We humans don't just sit around and watch things all day. We do things: solve problems by taking decisions and carrying them out. We move about and we manipulate our surroundings. (Sure, some days we stay in bed almost all day, but most of the rest of the time we are active in one way or another.) Our intelligence evolved to help us survive in a hostile environment, and doing that meant both reacting to the world and planning complicated sequences of actions, as well as adapting to changing circumstances. Pattern recognition - identifying objects and faces, understanding speech and so on - is an important component of intelligence, but should really be thought of as one part of a complete system which is constantly working on figuring out what to do next. Trying to invent artificial intelligence while only focusing on pattern recognition is like trying to invent the car while only focusing on the wheels.

In order to build a complete artificial intelligence we therefore need to build a system that takes actions in some kind of environment. How can we do this? Perhaps the most obvious idea is to embody artificial intelligence in robots. And indeed, robotics has shown us how even the most mundane tasks, such as walking in terrain or grabbing strangely shaped objects, are really rather hard to accomplish for robots. In the eighties, robotics research largely refocused on these kind of "simple" problems, which led to progress in applications as well as a better understanding of what intelligence is all about. The last few decades of progress in robotics has fed into the development of self-driving cars, which is likely to become one of the areas where AI technology will revolutionize society in the near future.

Once upon a time, when I thought I was a roboticist of sorts. The carlike robot in the foreground was supposed to learn to independently drive around the University of Essex campus (in the background). I think it's fair to say we underestimated the problem. (From left to right: me (now at NYU), Renzo de Nardi (Google), Simon Lucas (Essex), Richard Newcombe (Facebook/Oculus), Hugo Marques (ETH Zurich).)

Now, working with robots clearly has its downsides. Robots are expensive, complex and slow. When I started my PhD, my plan was to build robot software that would learn evolutionarily from its mistakes in order to develop increasingly complex and general intelligence. But I soon realized that in order for my robots to learn from their experiences, they would have to attempt each task thousands of times, with each attempt maybe taking a few minutes. This meant that even a simple experiment would take several days - even if the robot would not break down (it usually would) or start behaving differently as the batteries depleted or motors warmed up. In order to learn any more complex intelligence I would have to build an excessively complex (and expensive) robot with advanced sensors and actuators, further increasing the risk of breakdown. I also would have to develop some very complex environments where complex skills could be learned. This all adds up, and quickly becomes unmanageable. Problems such as these is why the field of evolutionary robotics has not scaled up to evolve more complex intelligence.
The basic idea of evolutionary robotics: use Darwinian evolution implemented as a computer algorithm to teach robot AIs (for example neural networks) to solve tasks. Image taken from a recent survey paper.

I was too ambitious and impatient for that. I wanted to create complex intelligence that could learn from experience. So I turned to video games.

Games and artificial intelligence have a long history together. Even since before artificial intelligence was recognized as a field, early pioneers of computer science wrote game-playing programs because they wanted to test whether computers could solve tasks that seemed to require "intelligence". Alan Turing, arguably the principal inventor of computer science, (re)invented the Minimax algorithm and used it to play Chess. (As no computer had been built yet, he performed the calculations himself using pen and paper.) Arthur Samuel was the first to invent the form of machine learning that is now called reinforcement learning; he used it in a program that learned to play Checkers by playing against itself. Much later, IBM's Deep Blue computer famously won against the reigning grandmaster of Chess, Gary Kasparov, in a much-publicized 1997 event. Currently, many researchers around the world work on developing better software for playing the board game Go, where the best software is still no match for the best humans.

Arthur Samuel in 1957, playing checkers against a mainframe computer the size of a room and thousands of times less computer power than your phone. The computer won. 

Classic board game such as Chess, Checkers and Go are nice and easy to work with as they are very simple to model in code and can be simulated extremely fast - you could easily make millions of moves per second on a modern computer - which is indispensable for many AI techniques. Also, they seem to require thinking to play well, and have the property that they take "a minute to learn, but a lifetime to master". It is indeed the case that games have a lot to do with learning, and good games are able to constantly teach us more about how to play them. Indeed, to some extent the fun in playing a game consists in learning it and when there is nothing more to learn we largely stop enjoying them. This suggests that better-designed games are also better benchmarks for artificial intelligence. However, judging from the fact that now have (relatively simple) computer programs that can play Chess better than any human, it is clear that you don't need to be truly, generally intelligent to play such games well. When you think about it, they exercise only a very narrow range of human thinking skills; it's all about turn-based movements on a discrete grid of a few pieces with very well-defined, deterministic behavior.

Deep Blue beats Kasparov in Chess.


But, despite what your grandfather might want you to believe, there's more to games than classical board games. In addition to all kinds of modern boundary-pushing board games, card games and role-playing games, there's also video games. Video games owe their massive popularity at least partly to that they engage multiple senses and multiple cognitive skills. Take a game such as Super Mario Bros. It requires you not only to have quick reactions, visual understanding and motoric coordination, but also to plan a path through the level, decide about tradeoffs between various path options (which include different risks and rewards), predict the future position and state of enemies and other characters of the level, predict the physics of your own running and jumping, and balance the demands of limited time to finish the level with multiple goals. Other games introduce demands of handling incomplete information (e.g. StarCraft), understanding narrative (e.g. Skyrim), or very long-term planning (e.g. Civilization).

A very simple car racing game I hacked up for a bunch of experiments that went into my PhD thesis. We also used it for a competition.

On top of this, video games are run inside controllable environments in a computer and many (though not all) video games can be sped up to many times the original speed. It is simple and cheap to get started, and experiments can be run many thousand times in quick succession, allowing the use of learning algorithms.

After the first year, the Simulated Car Racing Competition switched to the TORCS racing game, which is more advanced. It also undeniably looks better.

So it is not surprising that AI researchers have increasingly turned to video games as benchmarks recently. Researchers such as myself have adapted a number of video games to function as AI benchmarks. In many cases we have organized competitions where researchers can submit their best game-playing AIs and test them against the best that other researchers can produce; having recurring competitions based on the same game allows competitors to refine their approaches and methods, hoping to win next year. Games for which we have run such competitions include Super Mario Bros (paper), StarCraft (paper), the TORCS racing game (paper), Ms Pac-Man (paper), a generic Street Fighter-style figthing game (paper), Angry Birds (paper), Unreal Tournament (paper) and several others. In most of these competitions, we have seen performance of the winning AI player improve every time the competition is run. These competitions play an important role in catalyzing research in the community, and every year many papers are published where the competition software is used for benchmarking some new AI method. Thus, we advance AI through game-based competitions.

There's a problem with the picture I just painted. Can you spot it?

That's right. Game specificity. The problem is that improving how well an artificial intelligence plays a particular game is not necessarily helping us improve artificial intelligence in general. It's true that in most of the game-based competitions mentioned above we have seen the submitted AIs get better every time the competition ran. But in most cases, the improvements were not because of better AI algorithms, but because of even more ingenious ways of using these algorithms for the particular problems. Sometimes this meant relegating the AI to a more peripheral role. For example, in the car racing competition the first years were dominated by AIs that used evolutionary algorithms to train a neural network to keep the car on the track. In later years, most of the best submissions used hand-crafted "dumb" methods to keep the car on the track, but used learning algorithms to learn the shape of the track to adapt the driving. This is a clever solution to a very specific engineering problem but says very little about intelligence in general.
The game Ms. Pac-Man was used in a successful recent AI competition, which saw academic researchers, students and enthusiasts from around the world submit their AI Pac-Man players. The challenge of playing Pac-Man spurred researchers to invent new versions of state-of-the-art algorithms to try to get ahead. Unfortunately, the best computer players so far are no better than an intermediate human.
In order to make sure that what such a competition measures is anything approaching actual intelligence, we need to recast the problem. To do this, it's a great idea to define what it is we want to measure: general intelligence. Shane Legg and Marcus Hutter have proposed a very useful definition of intelligence, which is roughly the average performance of an agent on all possible problems. (In their original formulation, each problem's contribution to the average is weighed by its simplicity, but let's disregard that for now.) Obviously, testing an AI on all possible problems is not an option, as there are infinitely many problems. But maybe we could test our AI on just a sizable number of diverse problems? For example on a number of different video games?

The first thing that comes to mind here is to just to take a bunch of existing games for some game console, preferably one that could be easily emulated and sped up to many times real time speed, and build an AI benchmark on them. This is what the Arcade Learning Environment (ALE) does. ALE lets you test your AI on more than a hundred games released for 70s vintage Atari 2600 console. The AI agents get feeds of the screen at pixel level, and have to respond with a joystick command. ALE has been used in a number of experiments, including those by the original developers of the framework. Perhaps most famously, Google Deep Mind published a paper in Nature showing how they could learn to play several of the games with superhuman skill using deep learning (Q-learning on a deep convolutional network).

The Atari 2600 video game console. Note the 1977-vintage graphics of the Combat game on the TV in the background. Note, also, the 1977-vintage wooden panel. Your current game console probably suffers from a lack of wooden panels. Not in picture: 128 bytes of memory. That's a few million times less than your cell phone.

ALE is an excellent AI benchmark, but has a key limitation. The problem with using Atari 2600 games is that there is only a finite number of them, and developing new games is a trick process. The Atari 2600 is notoriously hard to program, and the hardware limitations of the console tightly constrain what sort of games can be implemented. More importantly, all of the existing games are known and available to everyone. This makes it possible to tune your AI to each particular game. Not only to train your AI for each game (DeepMind's results depend on playing each game many thousands of times to train on it) but to tune your whole system to work better on the games you know you will train on.

Can we do better than this? Yes we can! If we want to approximate testing our AI on all possible problems, the best we can do is to test it on a number of unseen problems. That is, the designer of the AI should not know which problems it is being tested on before the test. At least, this was our reasoning when we designed the General Video Game Playing Competition.

"Frogger" for the Atari 2600.

The General Video Game Playing Competition (GVGAI) allows anyone to submit their best AI players to a special server, which will then use them to play ten games that no-one (except the competition organizers) have seen before. These games are of the type that you could find on home computers or in arcades in the early eighties; some of them are based on existing games such as Boulder Dash, Pac-Man, Space Invaders, Sokoban and Missile Command. The winner of the competition is the AI that plays these unseen games best. Therefore, it is impossible for the creator of the AI to tune their software to any particular game. Around 50 games are available for training your AI on, and every iteration of the competition increases this number as the testing games from the previous iteration become available to train on.

"Frogger" implemented in VGDL in the General Video Game Playing framework.

Now, 50 games is not such a large number; where do we get new games from? To start with, all the games are programmed in something called the Video Game Description Language (VGDL). This is a simple language we designed to to be able to write games in a compact and human-readable way, a bit like how HTML is used to write web pages. The language is designed explicitly to be able to encode classical arcade games; this means that the games are all based on the movement of and interaction sprites in two dimensions. But that goes for essentially all video games before Wolfenstein 3D. In any case, the simplicity of this language makes it very easy to write new games, either from scratch or as variations on existing games. (Incidentally, as an offshoot of this project we are exploring the use of VGDL as a prototyping tool for game developers.)

A simple version of the classic puzzle game Sokoban implemented in VGDL.

Even if it's simple to write new games, that doesn't solve the fundamental problem that someone has to write them, and design them first. For the GVG-AI competition to reach its full potential as a test of general AI, we need an endless supply of new games. For this, we need to generate them. We need software that can produce new games at the press of a button, and these need to be good games that are not only playable but also require genuine skill to win. (As a side effect, such games are likely to be enjoyable for humans.)

Boulder Dash implemented in VGDL in the General Video Game Playing framework.

I know, designing software that can design complete new games (that are also good in some sense) sounds quite hard. And it is. However, I and a couple of others have been working on this problem on and off for a couple of years, and I'm firmly convinced it is doable. Cameron Browne has already managed to build a complete generator for playable (and enjoyable) board games, and some of our recent work has focused on generating simple VGDL games, though there is much left to do. Also, it is clearly possible to generate parts of games, such as game levels; there has been plenty of research within the last five years on procedural content generation - the automatic generation of game content. Researchers have demonstrated that methods such as evolutionary algorithms, planning and answer set programming can automatically create levels, maps, stories, items and geometry, and basically any other content type for games. Now, the research challenges are to make these methods general (so that they work for all games, not just for a particular game) and more comprehensive, so that they can generate all aspects of a game including the rules. Most of the generative methods include some form of simulation of the games that are being generated, suggesting that the problems of game playing and game generation are intricately connected and should be considered together whenever possible.
Yavalath, a board game designed by computer program designed by Cameron Browne. Proof that computers can design complete games.
Once we have extended the General Video Game Playing Competition with automated game generation, we have a much better way of testing generic game-playing ability than we have ever had before. The software can of course also be used outside of the competition, providing a way to easily test the general intelligence of game-playing AI.

So far we have only talked about how to best test or evaluate the general intelligence of a computer program, not how to best create one. Well, this post is about why video games are essential for inventing AI, and I think that I have explained that pretty well: they can be used to fairly and accurately benchmark AIs. But for completeness, let us consider which are the most promising methods for creating AIs of this kind. As mentioned above, (deep) neural networks have recently attracted lots of attention because of some spectacular results in pattern recognition. I believe neural networks and similar pattern recognition methods will have an important role to play for evaluating game states and suggesting actions in various game states. In many cases, evolutionary algorithms are more suitable than gradient-based methods when training neural networks for games.

But intelligence can not only be pattern recognition. (This is for the same reason that behaviorism is not a complete account of human behavior: people don't just map stimuli to responses§, sometimes they also think.) Intelligence must also incorporate some aspect of planning, where future sequences of actions can be played out in simulation before deciding what to do. Recently an algorithm called Monte Carlo Tree Search, which simulates the consequences of long sequences of actions by doing statistics of random actions, has worked wonders on the board game Go. It has also done very well on GVGAI. Another family of algorithms that has recently shown great promise on game planning tasks is rolling horizon evolution. Here, evolutionary algorithms are used not for long-term learning, but for short-term action planning.

I think the next wave of advances in general video game-playing AIs will come from ingenious combinations of neural networks, evolution and tree search. And from algorithms inspired by these methods. The important thing is that both pattern recognition and planning will be necessary in various different capacities. As usual in research, we cannot predict what will work well in the future (otherwise it wouldn't be called research), but I bet that exploring various combinations of these method will inspire the invention of the next generation of AI algorithms.

A video game such as The Elder Scrolls V: Skyrim requires a wide variety of cognitive skills to play well.

Now, you might object that this is a very limited view of intelligence and AI. What about text recognition, listening comprehension, storytelling, bodily coordination, irony and romance? Our game-playing AIs can't do any of this, no matter if it can play all the arcade games in the world perfectly. To this I say: patience! None of these things are required for playing early arcade games, that is true. But as we master these games and move on to include other genres of games in our benchmark, such as role-playing games, adventure games, simulation games and social network games, many of these skills will be required to play well. As we gradually increase the diversity of games we include in our benchmark, we will also gradually increase the breadth of cognitive skills necessary to play well. Of course, our game-playing AIs will have to get more advanced to cope. Understanding language, images, stories, facial expression and humor will be necessary. And don't forget that closely coupled with the challenge of general video game playing is the challenge of general video game generation, where plenty of other types of intelligence will be necessary. I am convinced that video games (in general) challenges all forms of intelligence except perhaps those closely related to bodily movement, and therefore that video game (in general) is the best testbed for artificial intelligence. An AI that can play almost any video game and create a wide variety of video games is, by any reasonable standard, intelligent.

"But why, then, are not most AI researchers working on general video game playing and generation?"
To this I say: patience!

A game such as Civilization V requires a different, but just as wide, skillset to play well.

This blog post became pretty long - I had really intended it to be just a fraction of what it is. But there was so much to explain. In case you've read this far, you might very well have forgotten what I said in the beginning by now. So let me recapitulate:

It is crucial for artificial intelligence research to have good testbeds. Games are excellent AI testbeds because they pose a wide variety of challenges, similarly to robotics, and are highly engaging. But they are also simpler, cheaper and faster, permitting a lot of research that is not practically possible with robotics. Board games have been used in AI research since the field started, but in the last decade more and more researchers have moved to video games because they offer more diverse and relevant challenges. (They are also more fun.) Competitions play a big role in this. But putting too much effort into AI for a single game has limited value for AI in general. Therefore we created the General Video Game Playing Competition and its associated software framework. This is meant to be the most complete game-based benchmark for general intelligence. AIs are evaluated on playing not a single video game, but on multiple games which the AI designer has not seen before. It is likely that the next breakthroughs in general video game playing will come from a combination of neural networks, evolutionary algorithms and Monte Carlo tree search. Coupled with the challenge of playing these games, is the challenge of generating new games and new content for these games. The plan is to have an infinite supply of games to test AIs on. While playing and generating simple arcade games tests a large variety of cognitive capacities - more diverse than any other AI benchmark - we are not yet at the stage where we test all of intelligence. But there is no reason to think we would not get there, given the wide variety of intelligence that is needed to play and design modern video games.

If you want to know more about these topics, I've linked various blog posts, articles and books from the text above. Most of the research I'm doing (and that we do in the NYU Game Innovation Lab) is in some way connected to the overall project I've been describing here. I recently put together a little overview of research I've done in the past few years, with links to most of my recent papers. Many more papers can be found on my web page. It is also important to realize that most of this work has a dual purpose: to advance artificial intelligence and to make games more fun. Many of the technologies that we are developing are to a lesser or greater extent focused on improving games. It's important to note that there is still important work to be done in advancing AI for particular games. In another recent blog post, I tried to envision what video games would be like if we had actual artificial intelligence. A somewhat recent paper I co-wrote tries to outline the whole field of AI in games, but it's rather long and complex so I would advise you to read the blog post first. You can also peruse the proceedings of the Computational Intelligence and Games and Artificial Intelligence and Interactive Digital Entertainment conference series.

Finally, it's important to note that there is plenty of room to get involved in this line of research (I think of it all as an overarching meta-project) as there are many, many open research questions. So if you are not already doing research on this I think you should start working in this field. It's exciting, because it's the future. So what are you waiting for?