Evaluation function

[ErinvanderVeen]

An important part of the minimax algorithm is a good evaluation function. How should we go about this? I reckon we should look at existing literature to devise a good weighted function, and then use simulated annealing to get optimal weights. We can make the program play itself for this purpose.

[emwestin]

I would also assume that using existing literature is the way to go in order to get a good weighted function.
For example, in the paper "Performance and energy efficiency analysis of a Reversi player for FPGAs and General Purpose Processors" by J. Olivito et al., they describe the following metrics:

• Mobility. It refers to the number of legal moves of a player. This is a very important strategic concept of the game. A good player tries to maximize its mobility while limiting the opponent’s one. This usually forces the opponent to make undesirable movements."
• Stable discs. Stable discs are those that once they are placed, they cannot be flipped anymore. They are generally desirable because they allow dominating its surrounding squares and definitely contribute to the final score.
• Corners and x-squares. Corners are considered valuable squares because they are stable and they are the key to get more stable discs around the corner. On the contrary, xsquares (corners adjacent squares placed on the major diagonals) are generally undesirable squares because they facilitate the opponent to reach the corners. Note that this applies only when the corresponding corner is empty
• Number of discs. This metric evaluates end-game boards.

Some or all of these should probably be included in our evaluation function since they are often mentioned in the strategies used by professional players.

[ErinvanderVeen]

The development of a world class Othello program is also a good paper that lists a bunch of things we should consider in the evaluation function.