Exploring Monte Carlo Negotiation Search with Nontrivial Agreements

The application of automated negotiations to general game playing is a research area with far-reaching implications. Non-zero sum games can be used to model a wide variety of real-world scenarios and automated negotiation provides a framework for more realistically modeling the behavior of agents in these scenarios. A particular recent development in this space is the Monte Carlo Negotiation Search (MCNS) algorithm, which can negotiate to find valuable cooperative strategies for a wide array of games (such as those of the Game Description Language). However, MCNS only proposes agreements corresponding to individual sequences of moves without any higher-level notions of conditional or stateful strategy. Our work attempts to lift this restriction. We present two contributions: extensions to the MCNS algorithm to support more complex agreements and an agreement language for GDL games suitable for use with our algorithm. We also present the results of a preliminary experiment in which we use our algorithm to search for an optimal agreement for the iterated prisoners dilemma. We demonstrate significant improvement of our algorithm over random agreement sampling, although further work is required to more consistently produce optimal agreements.

Deep Reinforcement Learning for General Game Playing

General Game Playing agents are required to play games they have never seen before simply by looking at a formal description of the rules of the game at runtime. Previous successful agents have been based on search with generic heuristics, with almost no work done into using machine learning. Recent advances in deep reinforcement learning have shown it to be successful in some two-player zero-sum board games such as Chess and Go. This work applies deep reinforcement learning to General Game Playing, extending the AlphaZero algorithm and finds that it can provide competitive results.

General Game Playing with Imperfect Information

General Game Playing is a field which allows the researcher to investigate techniques that might eventually be used in an agent capable of Artificial General Intelligence. Game playing presents a controlled environment in which to evaluate AI techniques, and so we have seen an increase in interest in this field of research. Games of imperfect information offer the researcher an additional challenge in terms of complexity over games with perfect information. In this article, we look at imperfect-information games: their expression, their complexity, and the additional demands of their players. We consider the problems of working with imperfect information and introduce a technique called HyperPlay, for efficiently sampling very large information sets, and present a formalism together with pseudo code so that others may implement it. We examine the design choices for the technique, show its soundness and completeness then provide some experimental results and demonstrate the use of the technique in a variety of imperfect-information games, revealing its strengths, weaknesses, and its efficiency against randomly generating samples. Improving the technique, we present HyperPlay-II, capable of correctly valuing information-gathering moves. Again, we provide some experimental results and demonstrate the use of the new technique revealing its strengths, weaknesses and its limitations.