Learning to play the chess variant Crazyhouse with deep learning and domain knowledge

Makovec, Anei (author), Guid, Matej (mentor), Pirker, Johanna (co-mentor)

In the evolving landscape of game-playing algorithms, Crazyhouse's reintroduction of captured pieces presents a unique challenge that distinguishes it from traditional chess. In this thesis, we explore a hybrid approach that combines domain knowledge with neural network-based evaluations, aiming for an optimal balance of performance. Through rigorous experiments, including self-play, matchups against a variant of the known program, Go-deep experiments, and move score deviations, we present compelling evidence for the effectiveness of a weighted sum of evaluations from a traditional evaluation function and an AlphaZero-style neural network. Remarkably, in our experiments, the combination of 75% neural network and 25% traditional evaluation consistently emerged as the most effective choice. Furthermore, we introduce the use of Best-Change rates, previously associated with evaluation quality, in the context of Monte Carlo tree search-based algorithms. Our approach may hold promise beyond Crazyhouse, especially in domains where established heuristic knowledge has proven effective. Additionally, it provides a basis for potentially clarifying chessboard decisions - a significant departure from the complexity of neural network decision-making. The classical evaluation function provides interpretable domain knowledge, offering a potential avenue for understandable decision-making.

• Language: English
• Keywords: Crazyhouse, chess variants, neural networks, domain knowledge, Best-Change rates, Monte Carlo tree search
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-154972
• COBISS.SI-ID: 187474947
• Publication date in RUL: 12.03.2024

Secondary language

• Language: Slovenian
• Title: Učenje igranja različice šaha Norišnica z uporabo globokega učenja in domenskega znanja

Abstract

Na razvijajočem se področju algoritmov za igranje iger ponuja Norišnica s ponovno postavitvijo osvojenih figur na šahovnico edinstven izziv, ki jo ločuje od tradicionalnega šaha. V magistrskem delu raziskujemo hibridni pristop, ki združuje domensko znanje z ocenami na osnovi nevronskih mrež, s ciljem doseči optimalno ravnovesje igralne moči. Z natančnimi poskusi, vključno s samo-igranjem, dvoboji z različico znanega programa, poskusi Go-deep in poskusi odstopanja ocen potez, predstavljamo prepričljive dokaze o učinkovitosti utežene vsote ocen tradicionalne ocenjevalne funkcije in nevronske mreže v slogu AlphaZero. Presenetljivo se je kombinacija 75 % ocene nevronske mreže in 25 % tradicionalne ocene dosledno izkazala za najučinkovitejšo izbiro v vseh naših poskusih. Poleg tega uvajamo uporabo odstotkov Best-Change, ki so velikokrat povezani s kakovostjo ocene, v kontekstu algoritmov na osnovi drevesnega preiskovanja Monte Carlo. Naš pristop bi se lahko uspešno uporabil tudi na drugih področjih, še posebej tistih, kjer se je uveljavljeno hevristično znanje izkazalo kot učinkovito. Poleg tega naš pristop predstavlja osnovo za morebitno razjasnitev igralnih odločitev na šahovnici - pomemben odmik od zapletenega odločanja nevronske mreže. Klasična ocenjevalna funkcija namreč ponuja človeku razumljivo domensko znanje in nudi potencial za razumljive igralne odločitve.

• Keywords: Norišnica, šahovske različice, nevronske mreže, domensko znanje, odstotki Best-Change, drevesno preiskovanje Monte Carlo