Implement hyperparameter search

pydeeprecsys/rl/agents/actor_critic.py

@@ -1,5 +1,5 @@
 from pydeeprecsys.rl.agents.agent import ReinforcementLearning
-from typing import Any, List
+from typing import Any, List, Optional
 from pydeeprecsys.rl.experience_replay.experience_buffer import ExperienceReplayBuffer
 from pydeeprecsys.rl.experience_replay.buffer_parameters import (
     ExperienceReplayBufferParameters,
@@ -12,27 +12,34 @@ class ActorCriticAgent(ReinforcementLearning):
     """Policy estimator using a value estimator as a baseline.
     It's on-policy, for discrete action spaces, and episodic environments.
     This implementation uses stochastic policies.
-    TODO: could be a sub class of reinforces"""
+    TODO: could be a sub class of reinforce"""
 
     def __init__(
         self,
         n_actions: int,
         state_size: int,
         discount_factor: int = 0.99,
-        learning_rate=1e-3,
+        actor_hidden_layers: Optional[List[int]] = None,
+        critic_hidden_layers: Optional[List[int]] = None,
+        actor_learning_rate=1e-3,
+        critic_learning_rate=1e-3,
     ):
+        if not actor_hidden_layers:
+            actor_hidden_layers = [state_size * 2, state_size * 2]
+        if not critic_hidden_layers:
+            critic_hidden_layers = [state_size * 2, int(state_size / 2)]
         self.episode_count = 0
         self.value_estimator = ValueEstimator(
             state_size,
-            [state_size * 2, int(state_size / 2)],
+            critic_hidden_layers,
             1,
-            learning_rate=learning_rate,
+            learning_rate=critic_learning_rate,
         )
         self.policy_estimator = PolicyEstimator(
             state_size,
-            [state_size * 2, state_size * 2],
+            actor_hidden_layers,
             n_actions,
-            learning_rate=learning_rate,
+            learning_rate=actor_learning_rate,
         )
         self.discount_factor = discount_factor
         # starts the buffer

pydeeprecsys/rl/agents/rainbow.py

@@ -1,5 +1,5 @@
 from numpy.random import RandomState
-from typing import Any, Optional
+from typing import Any, Optional, List
 from numpy import arange
 from copy import deepcopy
 from pydeeprecsys.rl.neural_networks.dueling import DuelingDDQN
@@ -34,18 +34,20 @@ def __init__(
         batch_size: int = 32,
         noise_sigma: float = 0.017,
         discount_factor: float = 0.99,
-        learning_rate: float = 0.99,
+        learning_rate: float = 0.0001,
+        hidden_layers: List[int] = None,
         random_state: RandomState = RandomState(),
         statistics: Optional[LearningStatistics] = None,
     ):
 
         self.network = DuelingDDQN(
-            input_size,
-            output_size,
-            learning_rate,
-            noise_sigma,
-            discount_factor,
+            n_input=input_size,
+            n_output=output_size,
+            learning_rate=learning_rate,
+            noise_sigma=noise_sigma,
+            discount_factor=discount_factor,
             statistics=statistics,
+            hidden_layers=hidden_layers,
         )
         self.target_network = deepcopy(self.network)
 

pydeeprecsys/rl/agents/reinforce.py

@@ -1,6 +1,6 @@
 import numpy as np
 from pydeeprecsys.rl.agents.agent import ReinforcementLearning
-from typing import Any, List
+from typing import Any, List, Optional
 from pydeeprecsys.rl.experience_replay.experience_buffer import ExperienceReplayBuffer
 from pydeeprecsys.rl.experience_replay.buffer_parameters import (
     ExperienceReplayBufferParameters,
@@ -17,14 +17,16 @@ def __init__(
         self,
         n_actions: int,
         state_size: int,
+        hidden_layers: Optional[List[int]] = None,
         discount_factor: int = 0.99,  # a.k.a gamma
         learning_rate=1e-3,
     ):
         self.episode_count = 0
-
+        if not hidden_layers:
+            hidden_layers = [state_size * 2, state_size * 2]
         self.policy_estimator = PolicyEstimator(
             state_size,
-            [state_size * 2, state_size * 2],
+            hidden_layers,
             n_actions,
             learning_rate=learning_rate,
         )

pydeeprecsys/rl/manager.py

@@ -142,9 +142,9 @@ def hyperparameter_search(
         agent: type,
         params: dict,
         default_params: dict,
-        learning_statistics: LearningStatistics,
         episodes: int = 100,
         runs_per_combination: int = 3,
+        verbose: bool = True,
     ) -> dict:
         """Given an agent class, and a dictionary of hyperparameter names and values,
         will try all combinations, and return the mean reward of each combinatio
@@ -154,15 +154,27 @@ def hyperparameter_search(
             if len(p_value) < 2:
                 continue
             for value in p_value:
-                rl = agent(**{p_name: value, **default_params})
+                rl = agent(**{**default_params, p_name: value})
+                learning_statistics = LearningStatistics()
                 combination_key = f"{p_name}={value}"
                 for run in range(runs_per_combination):
-                    print(f"Testing combination {p_name}={value} round {run}")
-                    self.train(rl=rl, max_episodes=episodes, should_print=False)
+                    self.train(
+                        rl=rl,
+                        max_episodes=episodes,
+                        should_print=False,
+                        statistics=learning_statistics,
+                    )
                     combination_results[combination_key].append(
-                        learning_statistics.moving_rewards[-1]
+                        learning_statistics.moving_rewards.iloc[-1]
                     )
-                    print(f"result was {learning_statistics.moving_rewards[-1]}")
+                    if verbose:
+                        print(
+                            f"\rTested combination {p_name}={value} round {run} "
+                            f"result was {learning_statistics.moving_rewards.iloc[-1]}"
+                            "\t\t",
+                            end="",
+                        )
+
         return combination_results
 
 

pydeeprecsys/rl/neural_networks/dueling.py

@@ -19,30 +19,37 @@ def __init__(
         n_input: int,
         n_output: int,
         learning_rate: float,
+        hidden_layers: List[int] = None,
         noise_sigma: float = 0.17,
         discount_factor: float = 0.99,
         statistics: Optional[LearningStatistics] = None,
     ):
         super().__init__()
+        if not hidden_layers:
+            hidden_layers = [256, 256, 64, 64]
         self.discount_factor = discount_factor
-        self._build_network(n_input, n_output, noise_sigma)
+        self._build_network(n_input, n_output, noise_sigma, hidden_layers=hidden_layers)
         self.optimizer = Adam(self.parameters(), lr=learning_rate)
         self.statistics = statistics
 
-    def _build_network(self, n_input: int, n_output: int, noise_sigma: float):
+    def _build_network(
+        self, n_input: int, n_output: int, noise_sigma: float, hidden_layers: List[int]
+    ):
         """Builds the dueling network with noisy layers, the value
         subnet and the advantage subnet. TODO: add `.to_device()` to Modules"""
-        self.fully_connected_1 = Linear(n_input, 256, bias=True)
-        self.fully_connected_2 = NoisyLayer(256, 256, bias=True, sigma=noise_sigma)
+        assert len(hidden_layers) == 4
+        fc_1, fc_2, value_size, advantage_size = hidden_layers
+        self.fully_connected_1 = Linear(n_input, fc_1, bias=True)
+        self.fully_connected_2 = NoisyLayer(fc_1, fc_2, bias=True, sigma=noise_sigma)
         self.value_subnet = Sequential(
-            NoisyLayer(256, 64, bias=True, sigma=noise_sigma),
+            NoisyLayer(fc_2, value_size, bias=True, sigma=noise_sigma),
             ReLU(),
-            Linear(64, 1, bias=True),
+            Linear(value_size, 1, bias=True),
         )
         self.advantage_subnet = Sequential(
-            NoisyLayer(256, 64, bias=True, sigma=noise_sigma),
+            NoisyLayer(fc_2, advantage_size, bias=True, sigma=noise_sigma),
             ReLU(),
-            Linear(64, n_output, bias=True),
+            Linear(advantage_size, n_output, bias=True),
         )
 
     def forward(self, state):

pydeeprecsys/tests/unit/test_actor_critic.py

@@ -11,7 +11,11 @@ def test_reinforce_init():
 def test_reinforce_interaction():
     manager = CartpoleManager()
     agent = ActorCriticAgent(
-        n_actions=2, state_size=4, discount_factor=0.95, learning_rate=0.001
+        n_actions=2,
+        state_size=4,
+        discount_factor=0.95,
+        actor_learning_rate=0.001,
+        critic_learning_rate=0.001,
     )
     learning_statistics = LearningStatistics()
     manager.train(

pydeeprecsys/tests/unit/test_manager.py

@@ -1,6 +1,38 @@
 from pydeeprecsys.rl.manager import MovieLensFairnessManager
+from pydeeprecsys.rl.agents.reinforce import ReinforceAgent
 
 
 def test_movie_lens_manager():
     manager = MovieLensFairnessManager()
     assert manager.env is not None
+
+
+def test_hyperparameter_search():
+    manager = MovieLensFairnessManager()
+    agent = ReinforceAgent
+
+    default_params = {
+        "n_actions": manager.env.action_space.n,
+        "state_size": manager.env.observation_space.shape[0],
+        "hidden_layers": [64, 64],
+        "discount_factor": 0.95,
+        "learning_rate": 0.0001,
+    }
+
+    optimize_params = {
+        "hidden_layers": [[64, 64], [128, 128], [256, 256]],
+        "discount_factor": [0.9, 0.95, 0.99],
+        "learning_rate": [0.00001, 0.0001, 0.001],
+    }
+
+    results = manager.hyperparameter_search(
+        agent=agent,
+        runs_per_combination=2,
+        episodes=10,
+        params=optimize_params,
+        default_params=default_params,
+    )
+
+    assert results is not None
+    assert len(results.items()) == 9
+    assert len(results["discount_factor=0.9"]) == 2