Training an AI Player to play Pong Using the GTM

We extend the work of [McGlinchey 2003], in which the author trained an AI player to play Pong from game observation data recorded from games played by humans. The data trained a Self Organising Map (SOM) , and it was found that the AI player played Pong with a human style of play. However one of the drawback of using the SOM was that the movement of the bat was jerky, due to quantisation of vectors. The author had to applying smoothing to the AI player’s bat to make the movement more realistic. It was also found that the AI players were easy to beat. In this paper we train the AI player using Generative Topographic Mapping (GTM) and we show that using the mean of the conditional density to estimate the bat’s position is better than using the mode.