Rinascimento: Playing Splendor-Like Games With Event-Value Functions

Abstract

In the realm of games research, artificial general intelligence algorithms often use score as the main reward signal for learning or playing actions. However, this has shown its limitations in scenarios where the rewards are very rare or absent until the end of the game. The problem is even more severe when the computational budget available is limited. This article proposes a new approach based on event logging: the game state triggers an event every time one of its features changes. These events are processed by an event-value function (EF) that assigns a value to a single action or a sequence. Experiments show that this approach can mitigate the problem of scarce rewards and improve the artificial intelligence performance compared with both the point-based heuristics and state-value functions. Furthermore, this represents a step forward in a finer control of the strategy adopted by the artificial agent, by describing a much richer and controllable behavioral space through EFs. Tuned EFs are able to neatly synthesize the relevance of the events in the game. Agents using an EF are also more robust when playing games with several opponents.