next up previous contents
Next: Specific Opponent Modeling Up: Opponent Modeling Previous: Opponent Modeling

Action Classification

The limited representational power of the original network was also identified as a possible bottleneck in improving performance. Previously, BPP merged the actions PASS and CALL, and again BET and RAISE. This was possible because the paired playing actions are mutually exclusive. However, it is unlikely that given the same winning probability and pot odds each dual is as good a choice as the other. Therefore, the actions were separated and the OPP Action node was refined to have four distinct classes ( RAISE, BET, CALL, PASS) instead of the original two. This change resulted in four separate conditional probability vectors and provided a more descriptive means of representing an opponent's betting behaviour.

Jason R Carlton