Added Evolutionary Strategies Network and added more example scripts

rltorch/action_selector/IdentitySelector.py

@@ -0,0 +1,14 @@
+from .ArgMaxSelector import ArgMaxSelector
+import torch
+class IdentitySelector(ArgMaxSelector):
+    def __init__(self, model, action_size, device = None):
+        super(IdentitySelector, self).__init__(model, action_size, device = device)
+    # random_act is already implemented in ArgMaxSelector
+    def best_act(self, state):
+        with torch.no_grad():
+            if self.device is not None:
+                state = state.to(self.device)
+            action = self.model(state).squeeze(0).item()
+        return action
+    def act(self, state):
+        return self.best_act(state)

rltorch/action_selector/StochasticSelector.py

@@ -5,7 +5,7 @@ import rltorch
 from rltorch.action_selector import ArgMaxSelector
 
 class StochasticSelector(ArgMaxSelector):
-    def __init__(self, model, action_size, memory, device = None):
+    def __init__(self, model, action_size, memory = None, device = None):
         super(StochasticSelector, self).__init__(model, action_size, device = device)
         self.model = model
         self.action_size = action_size

rltorch/action_selector/__init__.py

@@ -1,4 +1,5 @@
 from .ArgMaxSelector import * 
 from .EpsilonGreedySelector import * 
+from .IdentitySelector import * 
 from .RandomSelector import * 
 from .StochasticSelector import * 

rltorch/agents/A2CSingleAgent.py

@@ -1,5 +1,3 @@
-# Deprecated since the idea of the idea shouldn't work without having some sort of "mental model" of the environment
-
 from copy import deepcopy
 import numpy as np
 import torch

rltorch/agents/QEPAgent.py

@@ -0,0 +1,110 @@
+from copy import deepcopy
+import collections
+import torch
+from torch.distributions import Categorical
+import rltorch
+import rltorch.memory as M
+
+# Q-Evolutionary Policy Agent
+# Maximizes the policy with respect to the Q-Value function.
+# Since function is non-differentiabile, depends on the Evolutionary Strategy algorithm
+class QEPAgent:
+    def __init__(self, policy_net, value_net, memory, config, target_value_net = None, logger = None):
+        self.policy_net = policy_net
+        assert isinstance(self.policy_net, rltorch.network.ESNetwork)
+        self.policy_net.fitness = self.fitness
+        self.value_net = value_net
+        self.target_value_net = target_value_net
+        self.memory = memory
+        self.config = deepcopy(config)
+        self.logger = logger
+        self.policy_skip = 10
+
+    def fitness(self, policy_net, value_net, state_batch):
+      action_probabilities = policy_net(state_batch)
+      distributions = list(map(Categorical, action_probabilities))
+      actions = torch.tensor([d.sample() for d in distributions])
+      
+      with torch.no_grad():
+        state_values = value_net(state_batch)
+      obtained_values = state_values.gather(1, actions.view(len(state_batch), 1)).squeeze(1)
+
+      return -obtained_values.mean().item()
+
+    def learn(self, logger = None):
+        if len(self.memory) < self.config['batch_size']:
+            return
+        
+        if (isinstance(self.memory, M.PrioritizedReplayMemory)):
+            weight_importance = self.config['prioritized_replay_weight_importance']
+            # If it's a scheduler then get the next value by calling next, otherwise just use it's value
+            beta = next(weight_importance) if isinstance(weight_importance, collections.Iterable) else weight_importance
+            minibatch = self.memory.sample(self.config['batch_size'], beta = beta)
+            state_batch, action_batch, reward_batch, next_state_batch, not_done_batch, importance_weights, batch_indexes = M.zip_batch(minibatch, priority = True)
+        else:
+            minibatch = self.memory.sample(self.config['batch_size'])
+            state_batch, action_batch, reward_batch, next_state_batch, not_done_batch = M.zip_batch(minibatch)
+        
+        # Send to their appropriate devices
+        state_batch = state_batch.to(self.value_net.device)
+        action_batch = action_batch.to(self.value_net.device)
+        reward_batch = reward_batch.to(self.value_net.device)
+        next_state_batch = next_state_batch.to(self.value_net.device)
+        not_done_batch = not_done_batch.to(self.value_net.device)
+
+        state_values = self.value_net(state_batch)
+        obtained_values = state_values.gather(1, action_batch.view(self.config['batch_size'], 1))
+
+        with torch.no_grad():
+            # Use the target net to produce action values for the next state
+            # and the regular net to select the action
+            # That way we decouple the value and action selecting processes (DOUBLE DQN)
+            not_done_size = not_done_batch.sum()
+            next_state_values = torch.zeros_like(state_values, device = self.value_net.device)
+            if self.target_value_net is not None:
+                next_state_values[not_done_batch] = self.target_value_net(next_state_batch[not_done_batch])
+                next_best_action = self.value_net(next_state_batch[not_done_batch]).argmax(1)
+            else:
+                next_state_values[not_done_batch] = self.value_net(next_state_batch[not_done_batch])
+                next_best_action = next_state_values[not_done_batch].argmax(1)
+
+            best_next_state_value = torch.zeros(self.config['batch_size'], device = self.value_net.device)
+            best_next_state_value[not_done_batch] = next_state_values[not_done_batch].gather(1, next_best_action.view((not_done_size, 1))).squeeze(1)
+            
+        expected_values = (reward_batch + (self.config['discount_rate'] * best_next_state_value)).unsqueeze(1)
+
+        if (isinstance(self.memory, M.PrioritizedReplayMemory)):
+            value_loss = (torch.as_tensor(importance_weights, device = self.value_net.device) * ((obtained_values - expected_values)**2).squeeze(1)).mean()
+        else:
+            value_loss = F.mse_loss(obtained_values, expected_values)
+        
+        if self.logger is not None:
+            self.logger.append("Loss/Value", value_loss.item())
+
+        self.value_net.zero_grad()
+        value_loss.backward()
+        self.value_net.clamp_gradients()
+        self.value_net.step()
+
+        if self.target_value_net is not None:
+            if 'target_sync_tau' in self.config:
+                self.target_value_net.partial_sync(self.config['target_sync_tau'])
+            else:
+                self.target_value_net.sync()
+
+        if (isinstance(self.memory, M.PrioritizedReplayMemory)):
+            td_error = (obtained_values - expected_values).detach().abs()
+            self.memory.update_priorities(batch_indexes, td_error)
+
+        ## Policy Training
+        if self.policy_skip > 0:
+          self.policy_skip -= 1
+          return
+        self.policy_skip = 10
+        if self.target_value_net is not None:
+          self.policy_net.calc_gradients(self.target_value_net, state_batch)
+        else:
+          self.policy_net.calc_gradients(self.value_net, state_batch)
+        self.policy_net.clamp_gradients()
+        self.policy_net.step()
+

rltorch/agents/__init__.py

@@ -1,4 +1,5 @@
 from .A2CSingleAgent import *
 from .DQNAgent import *
 from .PPOAgent import *
+from .QEPAgent import *
 from .REINFORCEAgent import *

rltorch/network/ESNetwork.py

@@ -0,0 +1,66 @@
+import numpy as np
+import torch
+from .Network import Network
+from copy import deepcopy
+
+class ESNetwork(Network):
+    """
+    Network that functions from the paper Evolutionary Strategies (https://arxiv.org/abs/1703.03864)
+    fitness_fun := model, *args -> fitness_value (float)
+    We wish to find a model that maximizes the fitness function
+    """
+    def __init__(self, model, optimizer, population_size, fitness_fn, config, sigma = 0.05, device = None, logger = None, name = ""):
+        super(ESNetwork, self).__init__(model, optimizer, config, device, logger, name)
+        self.population_size = population_size
+        self.fitness = fitness_fn
+        self.sigma = sigma
+
+    # We're not going to be calculating gradients in the traditional way
+    # So there's no need to waste computation time keeping track
+    def __call__(self, *args):
+        with torch.no_grad():
+            result = self.model(*args)
+        return result
+
+
+    def _generate_noise_dicts(self):
+        model_dict = self.model.state_dict()
+        white_noise_dict = {}
+        noise_dict = {}
+        for key in model_dict.keys():
+            white_noise_dict[key] = torch.randn(self.population_size, *model_dict[key].shape)
+            noise_dict[key] = self.sigma * white_noise_dict[key]
+        return white_noise_dict, noise_dict
+
+    def _generate_candidate_solutions(self, noise_dict):
+        model_dict = self.model.state_dict()
+        candidate_solutions = []
+        for i in range(self.population_size):
+            candidate_statedict = {}
+            for key in model_dict.keys():
+                candidate_statedict[key] = model_dict[key] + noise_dict[key][i]
+            candidate = deepcopy(self.model)
+            candidate.load_state_dict(candidate_statedict)
+            candidate_solutions.append(candidate)
+        return candidate_solutions
+
+    def calc_gradients(self, *args):
+        ## Generate Noise
+        white_noise_dict, noise_dict = self._generate_noise_dicts()
+        
+        ## Generate candidate solutions
+        candidate_solutions = self._generate_candidate_solutions(noise_dict)
+        
+        ## Calculate fitness then mean shift, scale
+        fitness_values = torch.tensor([self.fitness(x, *args) for x in candidate_solutions])
+        if self.logger is not None:
+            self.logger.append(self.name + "/" + "fitness_value", fitness_values.mean().item())
+        fitness_values = (fitness_values - fitness_values.mean()) / (fitness_values.std() + np.finfo('float').eps)
+
+        ## Insert adjustments into gradients slot
+        self.zero_grad()
+        for name, param in self.model.named_parameters():
+            if param.requires_grad:
+                noise_dim_n = len(white_noise_dict[name].shape)
+                dim = np.repeat(1, noise_dim_n - 1).tolist() if noise_dim_n > 0 else []
+                param.grad = (white_noise_dict[name] * fitness_values.float().reshape(self.population_size, *dim)).mean(0) / self.sigma

rltorch/network/__init__.py

@@ -1,3 +1,4 @@
+from .ESNetwork import *
 from .Network import *
 from .NoisyLinear import *
 from .TargetNetwork import *