Navigating the Landscape of Multiplayer Games to Probe the Drosophila of AI
share
Multiplayer games have a long history in being used as key testbeds for evaluation and training in artificial intelligence (AI), aptly referred to as the "Drosophila of AI". Traditionally, researchers have focused on using games to build strong AI agents that, e.g., achieve human-level performance. This progress, however, also requires a classification of how 'interesting' a game is for an artificial agent, which requires characterization of games and their topological landscape. Tackling this latter question not only facilitates an understanding of the characteristics of learnt AI agents in games, but can also help determine what game an AI should address next as part of its training. Here, we show how network measures applied to so-called response graphs of large-scale games enable the creation of a useful landscape of games, quantifying the relationships between games of widely varying sizes, characteristics, and complexities. We illustrate our findings in various domains, ranging from well-studied canonical games to significantly more complex empirical games capturing the performance of trained AI agents pitted against one another. Our results culminate in a demonstration of how one can leverage this information to automatically generate new and interesting games, including mixtures of empirical games synthesized from real world games.
READ FULL TEXT