credit Mitchell Riley: we can drop 65 parameters for about the same strength; also train for longer

The idea is that the current model is M_theta(board, ply) = "score for player 0", but we can instead create a different model M'_theta(board) = "score for player 0 assuming that it's player 0's move" and for a certain pair of functions sigma, normalise, recover M_theta(board, ply) = sigma(ply) M'_theta(normalise(board, ply)).

This pair of functions is sigma(ply) = +/- 1 (parity), and normalise being a conditional interchange of pieces on the board depending on the parity of the ply. The only things stopping this from being completely straightforward is that nn_train.py evaluates M'_theta itself so we need to pass sigma(ply) to the training loop.

6x6 is still untrained.

1 files changed, 2 insertions, 4 deletions

diff --git a/include/actions.h b/include/actions.h
index d96033c..748ea38 100644
--- a/include/actions.h
+++ b/include/actions.h
@@ -58,10 +58,8 @@ extern int8_t move_deltas[4];
 void actions_init(const uint8_t board_size);
 
 // Generate the legal actions. The move list is drawn from a static
-// per-search-level scratch buffer: one buffer per recursion level, so nested
-// generate calls at different levels do not clobber each other. `level' is the
-// recursion depth within the current search (0 at the root). The returned
-// pointer is valid until the next generate call at the same level.
+// per-search-level scratch buffer, per recursion level. The returned pointer is
+// valid until the next generate call at the same level.
 action_t *actions_generate_at(tak_state_p state, uint32_t *num_actions,
                               const uint32_t level);