Chennai Mathematical Institute

Seminars




MSc Thesis Defence
4:00 pm, Lecture Hall 4
Playing Mancala without Human Knowledge

R Siddarth
Chennai Mathematical Institute.
18-05-18


Abstract

Game playing is a popular area within the field of artificial intelligence. Most agents in literature have hand-crafted features and are often trained on datasets obtained from expert human play. We implement a self- play based algorithm using neural networks for policy estimation and Monte Carlo Tree Search for policy improvement, with no input human knowledge that learns to play Mancala. We evaluate our learning algorithm for the classic and a traditional South Indian version (Pallanguzhi) of the game of Mancala. Our work is compared with random and greedy baselines, as well as a minimax agent.