优化PPO自适应回充与泛化特征

2026-04-26 18:35:23 +08:00
parent 00b26af3ed
commit e99a224d86
2 changed files with 101 additions and 18 deletions
--- a/agent_ppo/workflow/train_workflow.py
+++ b/agent_ppo/workflow/train_workflow.py
@@ -158,7 +158,7 @@ class EpisodeRunner:
                            result_str = "WIN"
                    else:
                        if fm.battery <= 0 or remaining_charge <= 0:
-                            final_reward = -4.0 + 6.0 * cleaning_ratio
+                            final_reward = -fm.battery_fail_penalty() + 4.0 * cleaning_ratio
                            result_str = "BATTERY_FAIL"
                        elif fm.npc_danger or fm.nearest_npc_dist <= 1:
                            final_reward = -3.0 + 6.0 * cleaning_ratio