AI Games

Goals:

• Create a software to compete in RBC competition and beat other competitors
• Focussing on solving the problem of POMDP in RBC

Objective:

The aim is to create a working Chess bot that performs better than the top chess engines in RBC

Method:

LSTM Approach:

• One hot encoding the belief state for input to LSTM which has history of length h.
• Output from LSTM is probability distribution over chess board for the position of 3x3 grid to be chosen.
• 3x3 grid is sampled according to the output of LSTM which is then multiplied with this probability distribution(LSTM output) for the gradient to flow.
• For the loss part we have weighted mean absolute error of 3x3 grid at the position revealed by LSTM at time-step t and t-1 and then maximizing this loss to capture piece movement.
• After solving the problem of POMDP in RBC we look forward to integrate this with a RL based chess algorithm such as MuZero, AlphaZero.