Update zone_observation.py

zone_observation.py

@@ -143,6 +143,9 @@ class ForecastConfig:
     belief_update_radius: int   = 2      # spatial propagation radius (zone cells)
     belief_increase_rate: float = 0.30   # update magnitude when event confirmed
     belief_decrease_rate: float = 0.05   # update magnitude when event absent
+    belief_prior_weight:  float = 0.70   # weight on episode prior vs. zone signal when
+                                          # seeding zone_belief at reset:
+                                          # prior_w*prior + (1-prior_w)*signal
 
     # --- Economics ---
     alert_value:          float = 100.0  # reward for correct advisory issuance
@@ -154,6 +157,10 @@ class ForecastConfig:
     zone_visit_bonus:       float = 1.5   # raw bonus, first visit to a zone only
     unvisited_zone_penalty: float = 40.0  # raw penalty * (unvisited/active) at terminate
 
+    uncertainty_decay:         float = 0.70  # forecast_uncertainty[zone] *= this on inspection
+    info_gain_scale:           float = 5.0   # multiplier on belief info-gain in step reward
+    uncertainty_penalty_scale: float = 5.0   # multiplier on mean uncertainty at termination
+
     economic_randomization: bool  = False
     clean_episode_ratio:    float = 0.7
     event_spatial_correlation: float = 0.85
@@ -194,6 +201,20 @@ class ForecastConfig:
             raise ValueError(f"ForecastConfig: zone_visit_bonus must be >= 0")
         if self.unvisited_zone_penalty < 0:
             raise ValueError(f"ForecastConfig: unvisited_zone_penalty must be >= 0")
+        if not (0.0 <= self.belief_prior_weight <= 1.0):
+            raise ValueError(
+                f"ForecastConfig: belief_prior_weight={self.belief_prior_weight} "
+                f"must be in [0, 1]"
+            )
+        if not (0.0 <= self.uncertainty_decay <= 1.0):
+            raise ValueError(
+                f"ForecastConfig: uncertainty_decay={self.uncertainty_decay} "
+                f"must be in [0, 1]"
+            )
+        if self.info_gain_scale < 0:
+            raise ValueError(f"ForecastConfig: info_gain_scale must be >= 0")
+        if self.uncertainty_penalty_scale < 0:
+            raise ValueError(f"ForecastConfig: uncertainty_penalty_scale must be >= 0")
         if self.horizon_days < 1:
             raise ValueError(f"ForecastConfig: horizon_days must be >= 1")
         if self.n_zones < 1:
@@ -230,6 +251,10 @@ class ForecastConfig:
         self.inspection_cost     = float(_clip(self.inspection_cost,     0.01,  1_000.0))
         self.zone_visit_bonus       = float(_clip(self.zone_visit_bonus,       0.0, 1_000.0))
         self.unvisited_zone_penalty = float(_clip(self.unvisited_zone_penalty, 0.0, 10_000.0))
+        self.belief_prior_weight       = float(_clip(self.belief_prior_weight,       0.0, 1.0))
+        self.uncertainty_decay         = float(_clip(self.uncertainty_decay,         0.0, 1.0))
+        self.info_gain_scale           = float(_clip(self.info_gain_scale,           0.0, 1_000.0))
+        self.uncertainty_penalty_scale = float(_clip(self.uncertainty_penalty_scale, 0.0, 1_000.0))
         self.prior_belief        = float(_clip(self.prior_belief,        0.001, 0.999))
         self.belief_floor        = float(_clip(self.belief_floor,        0.001, 0.5))
         self.clean_episode_ratio = float(_clip(self.clean_episode_ratio, 0.0,   1.0))
@@ -263,12 +288,16 @@ class ForecastConfig:
             "belief_update_radius": self.belief_update_radius,
             "belief_increase_rate": self.belief_increase_rate,
             "belief_decrease_rate": self.belief_decrease_rate,
+            "belief_prior_weight": self.belief_prior_weight,
             "alert_value": self.alert_value,
             "false_alert_penalty": self.false_alert_penalty,
             "miss_penalty": self.miss_penalty,
             "inspection_cost": self.inspection_cost,
             "zone_visit_bonus": self.zone_visit_bonus,
             "unvisited_zone_penalty": self.unvisited_zone_penalty,
+            "uncertainty_decay": self.uncertainty_decay,
+            "info_gain_scale": self.info_gain_scale,
+            "uncertainty_penalty_scale": self.uncertainty_penalty_scale,
             "economic_randomization": self.economic_randomization,
             "clean_episode_ratio": self.clean_episode_ratio,
             "event_spatial_correlation": self.event_spatial_correlation,
@@ -1378,6 +1407,36 @@ if __name__ == "__main__":
         _assert(False, "ForecastConfig accepted invalid forecast_backend")
     except ValueError:
         pass
+
+    _assert(cfg2.belief_prior_weight == 0.70,       "belief_prior_weight default round-trip")
+    _assert(cfg2.uncertainty_decay == 0.70,         "uncertainty_decay default round-trip")
+    _assert(cfg2.info_gain_scale == 5.0,            "info_gain_scale default round-trip")
+    _assert(cfg2.uncertainty_penalty_scale == 5.0,  "uncertainty_penalty_scale default round-trip")
+    cfg_belief = ForecastConfig(
+        belief_prior_weight=0.35,
+        uncertainty_decay=0.5,
+        info_gain_scale=2.0,
+        uncertainty_penalty_scale=8.0,
+    )
+    cfg_belief_back = ForecastConfig.from_dict(cfg_belief.to_dict())
+    _assert(cfg_belief_back.belief_prior_weight == 0.35,
+            "non-default belief_prior_weight round-trip")
+    _assert(cfg_belief_back.uncertainty_decay == 0.5,
+            "non-default uncertainty_decay round-trip")
+    _assert(cfg_belief_back.info_gain_scale == 2.0,
+            "non-default info_gain_scale round-trip")
+    _assert(cfg_belief_back.uncertainty_penalty_scale == 8.0,
+            "non-default uncertainty_penalty_scale round-trip")
+    try:
+        ForecastConfig(belief_prior_weight=1.5)
+        _assert(False, "ForecastConfig accepted belief_prior_weight out of [0,1]")
+    except ValueError:
+        pass
+    try:
+        ForecastConfig(info_gain_scale=-1.0)
+        _assert(False, "ForecastConfig accepted negative info_gain_scale")
+    except ValueError:
+        pass
     print(f"  ForecastConfig  rational_threshold={cfg.rational_termination_threshold:.4f}"
           f"  belief_floor={cfg.belief_floor:.4f}  pipeline fields OK")