I regret to inform you that the book is almost over now. I know this is a little bit sad—for you, not for me. It took a lot of effort to write this book, and I’m happy that I’m basically done. It all started with my cats being rehomed, and ten chapters later, we have explored game design, definitions of intelligence, narrow and general artificial intelligence, automatic creativity, and games that learn who you are and what you want, among other things. In order to structure the closing comments, we will here revisit the three broad claims that I made at the beginning of the book and outline how the discussion in the book supports these claims.
Games are the future of AI. Games provide the best benchmarks for AI because of the way they are designed to challenge many different human cognitive abilities, as well as for their technical convenience and the availability of human data. We have only began to scratch the surface of game-based AI benchmarks.
Games, especially well-designed games, are fantastic AI benchmarks because they are designed to test our cognitive capacities; they are engaging to play partly because they are unusually pedagogical intelligence tests. We have seen how board games have been used as AI benchmarks since the very beginning of AI research, but also how landmark achievements such as Deep Blue and AlphaGo have made board games less relevant for future AI testing. To some extent, the classic board games are simply done. Attention is shifting to video games, which provide a different and in many ways much richer set of challenges; the diversity of video games means that all the broad cognitive abilities from Cattell-Horn-Carroll theory are challenged by some video games. Many video games, including real-time strategy games such as StarCraft, are also much harder than any of the classic board games for various reasons, such as large branching factors and hidden information. But it’s not enough to create agents that play individual video games well. If we want to create more general artificial intelligence—which most agree is a central goal of AI, despite the disagreement over what intelligence is—we need to create agents that can play any games we present them with. For this, we need benchmarks and competitions that reflect the real diversity of video games. To make this feasible, we will probably need to generate these games at least partly automatically.
AI is the future of games. We now have much more capable AI methods than just a few years ago, and we are rapidly learning how to best apply them to games. The potential roles of AI in games go far beyond providing skillful opponents. We need to adapt our ways of thinking about game design to fully harness the power of advanced AI algorithms and enable a new generation of AI-augmented games.
There is a sizeable gulf between the AI methods that are being researched in academia and the AI methods employed in most games. While there exist some sophisticated techniques for controlling NPCs in games, more general-purpose AI methods are largely absent from commercial game development. Compared to other industries, game development is curiously unaffected by the current surge of interest in artificial intelligence. This is to a large extent because most game genres build on design blueprints that were laid down decades ago, when effective AI on consumer hardware was more or less impossible. Games are therefore designed to not need AI. To change this and harness some of the possibilities modern AI brings, we need to rethink game design, starting with the roles that AI can play. There are many unexplored roles for AI beyond standard NPC control. One particularly prominent role for AI algorithms is that of generating game content. Procedural content generation has been a feature of some games for decades, but new methods based on, for example, evolutionary algorithms make much more wide-ranging and controllable content generation possible. Player modeling is another key use case for AI in games, and modeling players’ preferences and behavior makes it possible to adapt games to particular players, suggesting a future where games could continuously reinvent themselves in response to what players do. They may even invent content players do not realize they want yet. Tools such as player modeling and content generation can also be useful for designers and make game development easier and more accessible. All these methods are, however, dependent on progress in AI agents that can play games in general.
Games and AI for games help us understand intelligence. By studying how humans play and design games, we can understand how they think, and we can attempt to replicate this thinking by creating game-playing and game-designing AI agents. Game design is a cognitive science; it studies thinking—human thinking and machine thinking.
The fact that some games are easy for humans to play but hard for current algorithms, and vice versa, is an important source of information on our thinking. It tells us how our thinking differs from that of the algorithms we currently have and can inspire us to create new AI methods. But it’s not only in the context of game playing that game AI can inform us about thinking. Algorithms for designing games, or parts of games, can be seen as models of human creativity. Trying to create software that can design games will give us some idea of how human creative processes, currently badly understood, work. And the differences between human and machine design will tell us more about this process and give us ideas for further creative algorithms.
Finally, I’ll reiterate that all of the research topics discussed in this book feed into each other. AI for games and games for AI are not the same thing, but advances in one of these endeavors will enable advances in the other. And there is much left to do in this young research field with so many open research questions in every direction. It is also an inherently interdisciplinary research field, where computer scientists, cognitive scientists, designers, and humanities scholars who care about games can contribute.
Perhaps you would like to join us?