feat(config): add prediction_window_size

 - controls the number of timesteps to include when calling the policy/value model. Default to 7 because why not, and I think it's the number of frames that AlphaZero stacked..? 🤷

Author: justindujardin

libraries/mathy_python/mathy/agents/a3c/worker.py

@@ -273,7 +273,9 @@ def run_episode(self, episode_memory: EpisodeMemory) -> float:
             # before_rnn_state_h = selector.model.unwrapped.embedding.state_h.numpy()
             # before_rnn_state_c = selector.model.unwrapped.embedding.state_c.numpy()
 
-            window = episode_memory.to_window_observation(last_observation)
+            window = episode_memory.to_window_observation(
+                last_observation, window_size=self.args.prediction_window_size
+            )
             try:
                 action, value = selector.select(
                     last_state=env.state,

libraries/mathy_python/mathy/agents/base_config.py

@@ -20,6 +20,9 @@ class Config:
 
     # One of "batch" or "layer"
     normalization_style = "layer"
+    # The number of previous timesteps to pass in with the current one
+    # when making predictions.
+    prediction_window_size: int = 7
 
     # Whether to use the LSTM or non-reccurrent architecture
     use_lstm: bool = True
@@ -62,7 +65,7 @@ class Config:
     clip_grouping_control = True
 
     # Include the time/type environment features in the embeddings
-    use_env_features = False
+    use_env_features = True
 
     # Include the node values floating point features in the embeddings
     use_node_values = True