Athénan wins sixteen gold medals at the Computer Olympiad

Abstract

Unlike Alpha Zero-like algorithms (Silver et al., 2018), Athénan is based on the Descent framework (Cohen-Solal, 2020). Thus, during the training process, it uses a variant of Unbounded Minimax (Korf and Chickering, 1996) called Descent, instead of Monte Carlo Tree Search, to construct the partial game tree used to determine the best action to play and to collect data for learning. With Descent, at each move, the best sequences of moves are iteratively extended until terminal states. During evaluations, another variant of Unbounded Minimax is used. This variant contains in particular a generic solver and it chooses the safest action to decide between actions. Moreover, contrary to Alpha Zero, Athénan does not use a policy network, only a value network. The actions therefore do not need to be encoded. In addition, unlike the Alpha Zero paradigm, with Athénan all data generated during the searches to determine the best actions to play is used for learning. As a result, much more data is generated per match (Cohen-Solal and Cazenave, 2023), and thus the training is done more quickly and does not require a (massive) parallelization to give good results (contrary to Alpha Zero). Athénan can use end-of-game heuristic evaluations to improve its level of play, such as game score or game length (in order to win quickly and lose slowly). Further improvements are described in (Cohen-Solal, 2020).