TEM branch merger

README.md

@@ -104,7 +104,7 @@ pip install NeuralPlayground==0.0.5
 
 If you want to contribute to the project, get the latest development version
 from GitHub, and install it in editable mode, including the "dev" dependencies:
-
+#### Unix (Linux and macOS)
 ```bash
 git clone https://github.com/SainsburyWellcomeCentre/NeuralPlayground/ --single-branch
 cd NeuralPlayground

examples/agent_examples/whittington_2020_run.py

@@ -14,7 +14,7 @@
 from neuralplayground.experiments import Sargolini2006Data
 
 simulation_id = "TEM_custom_sim"
-save_path = os.path.join(os.getcwd(), "results_sim")
+save_path = os.path.join(os.getcwd(), "examples", "agent_examples", "results_sim")
 # save_path = os.path.join(os.getcwd(), "examples", "agent_examples", "trained_results")
 agent_class = Whittington2020
 env_class = BatchEnvironment
@@ -23,56 +23,56 @@
 params = parameters.parameters()
 full_agent_params = params.copy()
 
+# Set the x and y limits for the arena
 arena_x_limits = [
-    [-5, 5],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-5, 5],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-5, 5],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-5, 5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
 ]
 arena_y_limits = [
-    [-5, 5],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-5, 5],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-5, 5],
-    [-4, 4],
-    [-5, 5],
-    [-6, 6],
-    [-5, 5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
+    [-2.5, 2.5],
 ]
 
-room_widths = [int(np.diff(arena_x_limits)[i]) for i in range(len(arena_x_limits))]
-room_depths = [int(np.diff(arena_y_limits)[i]) for i in range(len(arena_y_limits))]
-
+# Set parameters for the environment that generates observations
 discrete_env_params = {
     "environment_name": "DiscreteObject",
     "state_density": 1,
     "n_objects": params["n_x"],
-    "agent_step_size": 1,
+    "agent_step_size": 1,  # Note: this must be 1 / state density
     "use_behavioural_data": False,
     "data_path": None,
     "experiment_class": Sargolini2006Data,
 }
 
+# Set parameters for the batch environment
 env_params = {
     "environment_name": "BatchEnvironment",
     "batch_size": 16,
@@ -81,19 +81,33 @@
     "env_class": DiscreteObjectEnvironment,
     "arg_env_params": discrete_env_params,
 }
+
+# If behavioural data are used, set arena limits to those from Sargolini et al. 2006, reduce state density to 1/4
+state_densities = [discrete_env_params["state_density"] for _ in range(env_params["batch_size"])]
+if discrete_env_params["use_behavioural_data"]:
+    arena_x_limits = [[-50, 50] for _ in range(env_params["batch_size"])]
+    arena_y_limits = [[-50, 50] for _ in range(env_params["batch_size"])]
+    state_densities = [0.25] * env_params["batch_size"]
+
+room_widths = [int(np.diff(arena_x_limits)[i]) for i in range(env_params["batch_size"])]
+room_depths = [int(np.diff(arena_y_limits)[i]) for i in range(env_params["batch_size"])]
+
+# Set parameters for the agent
 agent_params = {
     "model_name": "Whittington2020",
+    "save_name": str(simulation_id)[4:],
     "params": full_agent_params,
     "batch_size": env_params["batch_size"],
     "room_widths": room_widths,
     "room_depths": room_depths,
-    "state_densities": [discrete_env_params["state_density"]] * env_params["batch_size"],
-    "use_behavioural_data": False,
+    "state_densities": state_densities,
+    "use_behavioural_data": discrete_env_params["use_behavioural_data"],
 }
 
 # Full model training consists of 20000 episodes
-training_loop_params = {"n_episode": 10, "params": full_agent_params}
+training_loop_params = {"n_episode": 5000, "params": full_agent_params, "random_state": False, "custom_state": [0.0, 0.0]}
 
+# Create the training simulation object
 sim = SingleSim(
     simulation_id=simulation_id,
     agent_class=agent_class,
@@ -104,7 +118,7 @@
     training_loop_params=training_loop_params,
 )
 
-# print(sim)
+# Run the simulation
 print("Running sim...")
 sim.run_sim(save_path)
 print("Sim finished.")

examples/comparisons_examples/comparison_board.ipynb

@@ -783,7 +783,7 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3.10.12 ('NPG-env')",
    "language": "python",
    "name": "python3"
   },

neuralplayground/agents/agent_core.py

@@ -82,7 +82,9 @@ def act(self, obs, policy_func=None):
 
         self.obs_history.append(obs)
         if len(self.obs_history) >= 1000:  # reset every 1000
-            self.obs_history.pop(0)
+            self.obs_history = [
+                obs,
+            ]
         if policy_func is not None:
             return policy_func(obs)