Update weather_forecast_env.py

weather_forecast_env.py

@@ -236,25 +236,25 @@ def _make_per_zone_forecasts(
     clean_ratio: float,
     spatial_corr: float = 0.85,
     event_plan: Optional[List[Dict[str, bool]]] = None,
+    flag_error_rate: float = 0.0,
 ) -> List[Any]:
     if event_plan is None:
-        event_plan = _episode_event_plan(
-            len(zone_ids), base_seed, clean_ratio, spatial_corr
-        )
+        event_plan = _episode_event_plan(len(zone_ids), base_seed, clean_ratio, spatial_corr)
+    err_rng = random.Random(base_seed + 1337)  # separate stream, never perturbs other draws
     forecasts = []
     for i, zid in enumerate(zone_ids):
-        ev = event_plan[i] if i < len(event_plan) else _zone_event_flags(
-            base_seed + i, clean_ratio
-        )
-        fc = make_synthetic_forecast_result(
-            zid,
-            horizon_days=horizon,
-            valid_time=valid_time,
-            seed=base_seed + i * 31,
-            drought=ev["drought"],
-            flood=ev["flood"],
-        )
-        forecasts.append(fc)
+        ev = event_plan[i] if i < len(event_plan) else _zone_event_flags(base_seed + i, clean_ratio)
+        drought, flood = ev['drought'], ev['flood']
+        if flag_error_rate > 0.0 and err_rng.random() < flag_error_rate:
+            if drought or flood:
+                drought = flood = False           # forecast misses a real event
+            elif err_rng.random() < 0.5:
+                drought, flood = True, False      # false drought alarm
+            else:
+                drought, flood = False, True      # false flood alarm
+        forecasts.append(make_synthetic_forecast_result(
+            zid, horizon_days=horizon, valid_time=valid_time,
+            seed=base_seed + i * 31, drought=drought, flood=flood))
     return forecasts
 
 
@@ -383,32 +383,20 @@ class WeatherForecastEnv(gym.Env):
         self._context_source: str = "synthetic"
 
         self._real_index = None
-        pkl_path = getattr(self.config, "real_data_pkl_path", None)
+        pkl_path = getattr(self.config, 'real_data_pkl_path', None)
         if pkl_path:
             if RealEpisodeIndex is None:
-                logger.warning(
-                    "WeatherForecastEnv: config.real_data_pkl_path=%r set, "
-                    "but real_episode_sampler.py is not importable -- "
-                    "training will remain fully synthetic.",
-                    pkl_path,
+                raise RuntimeError(
+                    f'real_data_pkl_path={pkl_path!r} is set but real_episode_sampler.py '
+                    'is not importable; refusing to fall back to synthetic training.'
                 )
-            else:
-                try:
-                    self._real_index = RealEpisodeIndex(pkl_path)
-                    logger.info(
-                        "WeatherForecastEnv: real-data training sampling "
-                        "ENABLED from %s (%d eligible dates, "
-                        "real_data_ratio=%.2f, inject_noise=%s)",
-                        pkl_path, self._real_index.n_eligible_dates,
-                        self.config.real_data_ratio, self.config.inject_noise,
-                    )
-                except Exception as e:
-                    logger.error(
-                        "WeatherForecastEnv: failed to load "
-                        "real_data_pkl_path=%r (%s) -- training will "
-                        "remain fully synthetic.", pkl_path, e,
-                    )
-                    self._real_index = None
+            self._real_index = RealEpisodeIndex(pkl_path)  # load errors propagate
+            logger.info(
+                'WeatherForecastEnv: real-data sampling ENABLED from %s '
+                '(%d eligible dates, real_data_ratio=%.2f, inject_noise=%s)',
+                pkl_path, self._real_index.n_eligible_dates,
+                self.config.real_data_ratio, self.config.inject_noise,
+            )
 
         logger.info(
             "WeatherForecastEnv: n_zones=%d  horizon=%dd  max_steps=%d  "
@@ -567,6 +555,7 @@ class WeatherForecastEnv(gym.Env):
                 self.config.clean_episode_ratio,
                 spatial_corr=rho,
                 event_plan=plan,
+                flag_error_rate=float(getattr(self.config, 'forecast_flag_error_rate', 0.0)),
             )
             self._zone_obs = _make_per_zone_obs(
                 self._zone_ids, effective_seed, self.config.clean_episode_ratio,