Towards artificially playing the game of double pong Combining learning and search algorithms

Abstract

This thesis work investigates ways of enhancing search methods used in reinforcement learning by utilizing neural networks. The environment in which the methods are tested is the classical game of pong. Two primary networks were used called Deep value network (DVN) and Fail-state network (F-Network). The first network aids the search by estimating state-values, the second network is used to detect search paths that lead to certain losses. Regarding the search methods, two algorithms were implemented, Random rollouts and Monte Carlo tree search (MCTS). It was concluded that the combination of search together with DVN drastically outperforms plain search methods, especially in environments where deep searches are unfeasible and CPU resources are restricted. The F-Network did not show any promising results in our study, however, possible improvements are discussed.