Abstract

Even with advents in deep reinforcement learning (RL), multi-agent gameplaying continues to be a difficult research challenge in Artificial Intelligence (AI)[8],[9]. This is especially true for games where the decision space is extremely large, as often seen in strategic board games like Risk and Monopoly [1],[15],[12]. The environment also contains relevant and irrelevant elements, making true generalization challenging. Beyond extensive self-play, it is not clear if optimal strategies can be learned by an agent in the context of specific games. Another problem that has not been discussed as much in the real world but that is important is the ‘applicability’or mapping of such games to actual strategic situations. Each game is ultimately an abstraction, and even in games like Monopoly, the only sources of variance are due to stochasticity (such as die rolls) or due to decisions made by other agents. In contrast, real-world environments can, and do, change in complex ways. We cannot anticipate, or plan for, every change in the real world. Real worlds are therefore ‘open’, in that some assertions can be made with certainty while the vast majority of facts are unknown [6]. More specifically, an open world refers to a world (or an environment) that is not fully parameterized and is open to unanticipated changes [6]. As noted above, the real world (as we know it) is the best example. Unexpected situations can (and do) occur, and while humans do not necessarily react ‘optimally’to such situations [14], they do (in many cases) react robustly. In other words, humans are capable of detecting, reacting, and adapting to unanticipated changes that they have not witnessed …

Date

2021

Authors

M Kejrivwal, S Thomas

Journal

Academia Letters

Pages

2