next up previous contents
Next: Comparison with alpha-beta Up: Comments about selective sampling Previous: Advantages

Simulation trade-offs

As encouraging as the simulation results have been, there is still room for improvement. Simulation can reveal information that results in better betting decisions and refine the estimate of the evaluation function. Yet, a simulation contains high variance and it is unclear how the other components of the program impact on its performance.

As was discussed in the previous section, simulation can recover from lack of knowledge in the evaluation function, produce complex game strategies and refine the estimated EV of the actions. This feature raises the question: how much knowledge does simulation require? Every time a new card is dealt in a trial, the hand strength and hand potential for active hands in the trial are calculated. For each betting action to be played, a PT generation function is called to generate an action for the opponents and Loki-2. Thus, during the simulation every trial is an accurate representation of a real game. However, each trial is a time-consuming process. In fact, one can exchange the accuracy of each trial for the number of trials performed in real-time. This tradeoff was explored by replacing the PT generation function betting strategy with an ``always call'' betting strategy inside the simulation. An ``always call'' betting strategy is probably the simplest betting strategy that can be provided to the simulation; it always returns call as all of the players' actions and all of Loki's subsequent actions. Therefore, in an ``always call'' simulation, there is no further betting after Loki's first decision, which is predetermined to be either call or raise, and every trial only consists of dealing all the cards and determining the hand that takes the pot. ``Always-call'' simulation-based Loki-2 wins against Loki-1 by a healthy margin (0.057 sb/hand) and runs 2.5 times faster than Loki-2 using the PT generation function. However it does not win as much as the full simulation-based Loki-2 does. Even though simple simulation is better than no simulation at all, knowledgeable simulation seems to provide better results.


next up previous contents
Next: Comparison with alpha-beta Up: Comments about selective sampling Previous: Advantages
Lourdes Pena
1999-09-10