fuse decision fusion

NN/models/dqn_agent.py

@@ -657,6 +657,24 @@ class DQNAgent:
             done: Whether episode is done
             is_extrema: Whether this is a local extrema sample (for specialized learning)
         """
+        # Validate states before storing experience
+        if state is None or next_state is None:
+            logger.debug("Skipping experience storage: None state provided")
+            return
+            
+        if isinstance(state, dict) and not state:
+            logger.debug("Skipping experience storage: empty state dictionary")
+            return
+            
+        if isinstance(next_state, dict) and not next_state:
+            logger.debug("Skipping experience storage: empty next_state dictionary")
+            return
+        
+        # Check if states are all zeros (invalid)
+        if hasattr(state, '__iter__') and all(f == 0 for f in np.array(state).flatten()):
+            logger.debug("Skipping experience storage: state is all zeros")
+            return
+            
         experience = (state, action, reward, next_state, done)
         
         # Always add to main memory
@@ -761,6 +779,15 @@ class DQNAgent:
             int: Action (0=BUY, 1=SELL) 
         """
         try:
+            # Validate state first - return early if empty/invalid/None
+            if state is None:
+                logger.warning("None state provided to act(), returning SELL action")
+                return 1  # SELL action (safe default)
+                
+            if isinstance(state, dict) and not state:
+                logger.warning("Empty state dictionary provided to act(), returning SELL action")
+                return 1  # SELL action (safe default)
+                
             # Use the DQNNetwork's act method for consistent behavior
             action_idx, confidence, action_probs = self.policy_net.act(state, explore=explore)
             
@@ -790,6 +817,14 @@ class DQNAgent:
     def act_with_confidence(self, state: np.ndarray, market_regime: str = 'trending') -> Tuple[int, float, List[float]]:
         """Choose action with confidence score adapted to market regime"""
         try:
+            # Validate state first - return early if empty/invalid/None
+            if state is None:
+                logger.warning("None state provided to act_with_confidence(), returning safe defaults")
+                return 1, 0.1, [0.0, 0.9, 0.1]  # SELL action with low confidence
+                
+            if isinstance(state, dict) and not state:
+                logger.warning("Empty state dictionary provided to act_with_confidence(), returning safe defaults")
+                return 1, 0.0, [0.0, 1.0]  # SELL action with zero confidence
             # Convert state to tensor if needed
             if isinstance(state, np.ndarray):
                 state_tensor = torch.FloatTensor(state)