RECONNAISSANCE BLIND CHESS

Registered Bots


Baseline Bots


Bot Description
trout Keeps track of a naive single board state with basic, semi-random sensing of unoccupied squares and makes moves with Stockfish.
attacker Senses randomly and tries to capture the opponent king with a classic 4-move "mate" or by attacking with a knight.
random Senses and moves randomly.
Oracle Keeps track of all the possible board states, sensing to minimize possible opponent states, and makes moves with Stockfish plus heuristics.
Marmot Uses a modified Monte Carlo counterfactual regret minimization algorithm for sensing and moving that leverages online outcome sampling and a heuristic state evaluation.
Zugzwang Keeps track of a single board state and uses rule-based catalogs to make sensing and moving decisions. The move catalog leverages a heavily modified version of Stockfish.
Bandit Uses perturbed response learning for sensing and movement, an approach that attempts to provoke opponent responses accounting for their intent.
DeterminedBot Uses a trained perfect-information approach to chess (Monte Carlo tree search guided by a neural network) coupled with a minimum-entropy sensing strategy. When acting under imperfect information, the algorithm samples over possible boards, computes a policy for each based on the trained tree search, and combines the set of resulting policies to choose a move.
ubuntu Uses a derived tree-search strategy that is guided by a neural network and learns to advantageously traverse from one set of possible boards to another to determine moves. Uses a minimum-entropy sensing strategy.