Experiments

The experimental design used to test the performance of the improved opponent modeling is the same as described in Section 4.3.1.

Refining action classifications appeared to have no significant effect upon performance, winning only $0.0202 \pm 0.0272$ units per game ( $t = 0.7442; p \le 0.025$) as seen in Figure 5. But when action refinement was combined with refined hand classes a significant improvement (over just implementing refined hands) followed ( $t = -4.0480; p \le 0.025$), winning $0.1061 \pm 0.0264$ units per game. This indicated that the performance gains achieved by refining the way that actions are classified is also highly dependent upon the level of hand class refinement performed.

  

Figure 5: Cumulative winnings of original BPP against players with refined action representations.