Fix bug in target Q-value for illegal actions

Fix DQN bug: set ILLEGAL_ACTION_LOGITS_PENALTY to a large negative number instead of 0.
google-deepmind · Aug 2, 2024 · f10eb08 · f10eb08
1 parent 1428a82
commit f10eb08
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/open_spiel/python/pytorch/dqn.py b/open_spiel/python/pytorch/dqn.py
@@ -30,7 +30,7 @@
     "Transition",
     "info_state action reward next_info_state is_final_step legal_actions_mask")
 
-ILLEGAL_ACTION_LOGITS_PENALTY = sys.float_info.min
+ILLEGAL_ACTION_LOGITS_PENALTY = torch.finfo(torch.float).min
 
 
 class SonnetLinear(nn.Module):