Activity: Talk or presentation › Scientific poster presentation
Description
Temporal credit assignment is a well-known problem in sequential game play. In this setting, players take turns making a move. It remains challenging to identify which were the decisive moves that lead to a win or cause a game to be lost. We introduce a novel approach that leverages the attention mechanism of transformers, specifically utilizing only the decoder architecture, to analyze moves in a cooperative or adversarial setting. The self-attention feature of transformers enables us to focus on varying parts of a game sequence and improve our understanding of how past actions influence future decisions and game outcomes. By training a simple transformer model on game play data, we examine the attention weights to determine the significance of each move at different stages of the game. Our analysis seeks to answer how to predict next moves and to identify crucial actions in winning strategies.