Tuesday, July 09, 2013

[gaptnwlz] Best response instead of minimax

Consider two computer programs playing a game, say chess, against each other, where both programs are given access to the other's evaluation function (and pruning function, etc.).  Instead of substituting one's own evaluation as a proxy, a guess, for how the opponent will evaluate a position, one can know for certain (modulo search depth).

We are straying from the assumption of Nash equilibrium.

There may be devilish details.  Is alpha-beta possible?

Probably want an API so the two programs can be linked without the need for interprocess or network communication.

Inspiration is testing and tuning a program by pitting many variations against each other.

Previous similar thoughts.

No comments :