I trained a player like Player 2.1 (race & contact networks, one-sided bearoff db), but with only five hidden nodes. I included the new inputs (shot hitting probability and expanded prime count).
The purpose here is to build a strategy that does something reasonably good - not optimal - but quick to calculate. We'll use it later in filtering possible moves in multiple-ply and rollout calculations to speed them up.
It ends up being better than pubEval, though not by much: it scores +0.115ppg and wins 54% of the games in a 10k-money game match.
I'll name it Player 2.3q, where the q denotes "quick".
The purpose here is to build a strategy that does something reasonably good - not optimal - but quick to calculate. We'll use it later in filtering possible moves in multiple-ply and rollout calculations to speed them up.
It ends up being better than pubEval, though not by much: it scores +0.115ppg and wins 54% of the games in a 10k-money game match.
I'll name it Player 2.3q, where the q denotes "quick".
Deprecated: I threw out the expanded prime count inputs in favor of Berliner-style prime inputs. Player 2.3q has been replaced by Player 2.4q, using the Berliner prime inputs and still 5 hidden nodes.
ReplyDelete