Imarisha: Reinforcement Learning Toolkit For AfricanBoard Games

EasyChair Preprint 2807

Abstract

Abstract—Popular board games go, shogi and chess, offer challenging test beds for research in planning and learning in reinforcement learning. However these environments are complex . The number of possible positions in a game of GO ranges from 170 10 ( 10 48 ) to 10 (5.310 ) . Similarly the number of positions in a game of chess ranges from 10 43 to 10 50 . With the current state of the art Algorithms, a massive amount of compute is needed to get robust performance in these environments. The Imarisha learning suite provides a set of diverse, two player learning board game environments that vary widely in complexity. This games majorly from the African cultures provides us with a rich variety of complexity, from Ajua to FANORONA. The learning suite offers environments of increasing complexity for RL research. Researchers can use the environments provided as stepping stones when testing the scalability and efficiency of their algorithms on harder and harder problems

Keyphrases: Deep Reinforcement Learning, Reinforcement Learning, african board game, african game, board game, board games, games, strategy board game

@booklet{EasyChair:2807,
  author    = {Eric Ondenyi},
  title     = {Imarisha: Reinforcement Learning Toolkit For AfricanBoard Games},
  howpublished = {EasyChair Preprint 2807},
  year      = {EasyChair, 2020}}