Descent wins five gold medals at the Computer Olympiad

Unlike AlphaZero-like algorithms (Silver et al., 2018), the Descent framework uses a variant of Unbounded Minimax (Korf and Chickering, 1996), 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. During training, at each move, the best sequences of moves are iteratively extended until terminal states. During evaluations, the safest action is chosen (after that the best sequences of moves are iteratively extended each until a leaf state is reached). Moreover, it also does not use a policy network, only a value network. The actions therefore do not need to be encoded. Unlike the AlphaZero paradigm, with Descent 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 game, and thus the training is done more quickly and does not require a (massive) parallelization to give good results (contrary to AlphaZero). It can use end-of-game heuristic evaluation to improve its level of play faster, such as game score or game length (in order to win quickly and lose slowly).