Beyond Minimax: Advanced Tree Search Algorithms and Machine Learning for Sequential Decision-Making — From Games to Supply Chain Optimization

This thesis develops a search-and-evaluate methodology for sequential decision-making on board games and tests whether it extends to supply-chain management. On Tic-Tac-Toe, Connect 4 and Othello, four search algorithms are implemented: Minimax, Alpha-Beta pruning, Monte Carlo Tree Search and Descent Minimax. A convolutional neural network, trained offline on millions of recorded Othello positions, then replaces the handcrafted evaluator in each of them. Benchmarked over several hundred games with Wilson confidence intervals, the network improves every algorithm, with win rates of 78–100% over the heuristic version. The Beer Game, a four-echelon supply-chain simulation, is then formalized as a Markov Decision Process. Minimax and Alpha-Beta do not transfer to its cooperative setting. Monte Carlo Tree Search transfers directly, and Descent transfers once adapted to uncertainty. The principle of a learned evaluator replacing a handcrafted one transfers fully. The thesis ends with a proposed architecture for this setting and leaves its empirical validation to future work.