To achieve general intelligence, agents will have to reason about interactions with other entities within their environment, some of which may also possess general intelligence capabilities. These other agents might have different goals, which may be aligned or in conflict with an agent's goals, which must be taken into account when deciding how to act.
Games provide an abstract and formal model of such environments: each player has a well-defined goal and rules to describe the effects of interactions among the players. For an agent to be successful in achieving its goal, it must anticipate, act, and react strategically to the behavior of the other players. Furthermore, there are many popular games that have been played and studied for thousands of years that have been widely-accepted as grand challenges for rational decision-making. As a result, games have been commonly used as benchmarks to demonstrate the potential for complex reasoning in the field of artificial intelligence since its inception.
The first successful achievements in playing these games at super-human level were attained with methods that relied on and exploited domain expertise that were designed manually (e.g. chess, checkers). In recent years, we have seen examples of general approaches that learn to play these games via self-play reinforcement learning (RL), as first demonstrated in Backgammon. By relying less on domain expertise, these approaches have generalized more easily and have been seen success in several challenging games such as Go, chess, shogi, Hex, and even complex computer games such as Atari, Doom, Dota 2, Starcraft, and Capture-the-Flag.
Successful applications have, in part, leveraged recent advances in deep reinforcement learning, networks architectures, tree search enhanced policy improvement, game theoretic modeling, learning on different time scales, and meta-learning. While progress is impressive, we believe we have just scratched the surface of what is capable, and much work remains to be done in order to truly understand the algorithms and reasoning processes within these environments.