Sunday, November 20, 2016

[djekhfwt] Hash size affects result

We modified the Stockfish chess engine to store a 64-bit key (up from 16 bits) in each transposition table entry (TTEntry).  Previously.  This effectively raises the hash key size to at least 76 bits, so there should be approximately no hash collisions.  Nevertheless, the size of the hash table affects the result of the computation to a given fixed depth.

The transposition table is not acting strictly as memoization.  If it were strictly memoization, then a smaller hash size would simply result in repeated computation, taking longer to compute to a given fixed depth, but the same answer.

Source code.

Logs demonstrating the issue.

Future: measure which memory size is stronger (for a given fixed depth) with a head-to-head match.

No comments :