Masinõpe k-ritta mängude õppimiseks
Abstract
Antud töö põhieesmärgiks oli uurida kui efektiivne ja mõistlik on kombineerida mitu
erinevat masinõppe meetodit, et treenida tehisintellekti k-ritta tüüpi mängudele. Need
meetodid on järgnevad: geneetiline algoritm, juhumetsad (koos otsustuspuudega) ning
Minimax algoritm. Eriliseks teeb sellise meetodi asjaolu, et kogu intelligents treenitakse ilma
inimese ekspert teadmisteta ning kõik vajaliku informatsiooni peab arvuti ise endale
omandama. The main objective of the thesis is to explore the viability of combination multiple
machine learning techniques in order to train Artificial Intelligence for k-in-a-row type games.
The techniques under observation are following:
- Random Forest
- Minimax Algorithm
- Genetic Algorithm
The main engine for training AI is Genetic Algorithm where a set of individuals are evolved
towards better playing computer intelligence. In the evaluation step, series of games are done
where individuals compete in series of games against each other – the results are recorded and
the evaluation score of the individuals are based on their performance in the games. During a
game, heuristic game tree search algorithm Minimax is used as player move advisor. Each of
the competing individuals has a Random Forest attached that is used as the heuristic function
in Minimax. The key idea of the training is to evolve as good Random Forests as possible. This
is achieved without any help of human expertise by using solely evolutionary training.