Automatic Tuning of a Chess Evaluation Function

I’ve taken a little detour over the past week to look at Population Based Incremental Tuning.  This was something which Thomas Petzke brought to my attention.  You can read his take here:

It’s an interesting approach.  After some investigation I’ve decide to add some form of automated tuning to Maverick’s chess position evaluation function.  I think PBIL seems to hold some promise.

  • Thomas Petzke

    I’m glad to give some motivation. It is still a lot of effort to get the tuning framework up and running, but more interesting than to play 30.000 games just to decide whether a single value should be increased or decreased a bit. A preliminary test shows that one of my PBIL tuned versions is definitely stronger than the one with the original (handpicked) weights.

    Currently I repeat the run to test whether it will converge to similar values and if not how different evolved end versions are in strength.