According to the results of self-play experiments,
a selective sampling simulation-based betting strategy for
Loki-2 significantly outperforms the static-evaluation based alternatives.
Similar to what has been seen with brute-force search in
games like chess, the effect of the simulation (search) amplifies the
quality of the evaluation function, allowing high performance to be
achieved without adding additional expert knowledge. Selective sampling uses the data
available about the game and the opponents to increase the quality of
the information obtained with each simulation run. Yet, the work on
selective-sampling simulation in poker is still in its early stages. The
knowledge component and selection methods have to be tuned with the algorithmic
component of the simulation, and the right balance between the different
simulation tradeoffs (cost per trial versus number of trials, random versus
systematic approach) has to be found.