added individual eps-greedy for VDN

algorithms/common.py

@@ -6,21 +6,6 @@ import torch
 import torch.nn as nn
 
 
-class BaseLearner:
-    def __init__(self, env, n_agents, lr):
-        self.env = env
-        self.n_agents = n_agents
-        self.lr = lr
-        self.device = 'cpu'
-
-    def to(self, device):
-        self.device = device
-        for attr, value in self.__dict__.items():
-            if isinstance(value, nn.Module):
-                value = value.to(self.device)
-        return self
-
-
 class Experience(NamedTuple):
     # can be use for a single (s_t, a, r s_{t+1}) tuple
     # or for a batch of tuples
@@ -29,6 +14,84 @@ class Experience(NamedTuple):
     action:           np.ndarray
     reward:           Union[float, np.ndarray]
     done  :           Union[bool, np.ndarray]
+    episode:          int = -1
+
+
+class BaseLearner:
+    def __init__(self, env, n_agents=1, train_every=('step', 4), n_grad_steps=1):
+        assert train_every[0] in ['step', 'episode'], 'train_every[0] must be one of ["step", "episode"]'
+        self.env = env
+        self.n_agents = n_agents
+        self.n_grad_steps = n_grad_steps
+        self.train_every = train_every
+        self.device = 'cpu'
+        self.n_updates = 0
+        self.step = 0
+        self.episode_step = 0
+        self.episode = 0
+        self.running_reward = deque(maxlen=5)
+
+    def to(self, device):
+        self.device = device
+        for attr, value in self.__dict__.items():
+            if isinstance(value, nn.Module):
+                value = value.to(self.device)
+        return self
+
+    def get_action(self, obs) -> Union[int, np.ndarray]:
+        pass
+
+    def on_new_experience(self, experience):
+        pass
+
+    def on_step_end(self, n_steps):
+        pass
+
+    def on_episode_end(self, n_steps):
+        pass
+
+    def train(self):
+        pass
+
+    def learn(self, n_steps):
+        train_type, train_freq = self.train_every
+        while self.step < n_steps:
+            obs, done = self.env.reset(), False
+            total_reward = 0
+            self.episode_step = 0
+            while not done:
+
+                action = self.get_action(obs)
+
+                next_obs, reward, done, info = self.env.step(action if not len(action) == 1 else action[0])
+
+                experience = Experience(observation=obs, next_observation=next_obs,
+                                        action=action, reward=reward,
+                                        done=done, episode=self.episode)  # do we really need to copy?
+                self.on_new_experience(experience)
+                # end of step routine
+                obs = next_obs
+                total_reward += reward
+                self.step += 1
+                self.episode_step += 1
+                self.on_step_end(n_steps)
+                if train_type == 'step' and (self.step % train_freq == 0):
+                    self.train()
+                    self.n_updates += 1
+            self.on_episode_end(n_steps)
+            if train_type == 'episode' and (self.episode % train_freq == 0):
+                self.train()
+                self.n_updates += 1
+
+            self.running_reward.append(total_reward)
+            self.episode += 1
+            try:
+                if self.step % 10 == 0:
+                    print(
+                        f'Step: {self.step} ({(self.step / n_steps) * 100:.2f}%)\tRunning reward: {sum(list(self.running_reward)) / len(self.running_reward):.2f}\t'
+                        f' eps: {self.eps:.4f}\tRunning loss: {sum(list(self.running_loss)) / len(self.running_loss):.4f}\tUpdates:{self.n_updates}')
+            except Exception as e:
+                pass
 
 
 class BaseBuffer:
@@ -60,7 +123,7 @@ def soft_update(local_model, target_model, tau):
 
 
 def mlp_maker(dims, flatten=False, activation='elu', activation_last='identity'):
-    activations = {'elu': nn.ELU, 'relu': nn.ReLU,
+    activations = {'elu': nn.ELU, 'relu': nn.ReLU, 'sigmoid': nn.Sigmoid,
                   'leaky_relu': nn.LeakyReLU, 'tanh': nn.Tanh,
                   'gelu': nn.GELU, 'identity': nn.Identity}
     layers = [('Flatten', nn.Flatten())] if flatten else []
@@ -71,7 +134,6 @@ def mlp_maker(dims, flatten=False, activation='elu', activation_last='identity')
     return nn.Sequential(OrderedDict(layers))
 
 
-
 class BaseDQN(nn.Module):
     def __init__(self, dims=[3*5*5, 64, 64, 9]):
         super(BaseDQN, self).__init__()

algorithms/q_learner.py

@@ -11,9 +11,9 @@ from algorithms.common import BaseLearner, BaseBuffer, soft_update, Experience
 
 class QLearner(BaseLearner):
     def __init__(self, q_net, target_q_net, env, buffer_size=1e5, target_update=3000, eps_end=0.05, n_agents=1,
-                 gamma=0.99, train_every_n_steps=4, n_grad_steps=1, tau=1.0, max_grad_norm=10, weight_decay=1e-2,
+                 gamma=0.99, train_every=('step', 4), n_grad_steps=1, tau=1.0, max_grad_norm=10, weight_decay=1e-2,
                  exploration_fraction=0.2, batch_size=64, lr=1e-4, reg_weight=0.0, eps_start=1):
-        super(QLearner, self).__init__(env, n_agents, lr)
+        super(QLearner, self).__init__(env, n_agents, train_every, n_grad_steps)
         self.q_net = q_net
         self.target_q_net = target_q_net
         self.target_q_net.eval()
@@ -26,11 +26,10 @@ class QLearner(BaseLearner):
         self.exploration_fraction = exploration_fraction
         self.batch_size = batch_size
         self.gamma = gamma
-        self.train_every_n_steps = train_every_n_steps
-        self.n_grad_steps = n_grad_steps
         self.tau = tau
         self.reg_weight = reg_weight
         self.weight_decay = weight_decay
+        self.lr = lr
         self.optimizer = torch.optim.AdamW(self.q_net.parameters(),
                                            lr=self.lr,
                                            weight_decay=self.weight_decay)
@@ -64,36 +63,14 @@ class QLearner(BaseLearner):
             action = np.array([self.env.action_space.sample() for _ in range(self.n_agents)])
         return action
 
-    def learn(self, n_steps):
-        step = 0
-        while step < n_steps:
-            obs, done = self.env.reset(), False
-            total_reward = 0
-            while not done:
+    def on_new_experience(self, experience):
+        self.buffer.add(experience)
 
-                action = self.get_action(obs)
-
-                next_obs, reward, done, info = self.env.step(action if not len(action) == 1 else action[0])
-
-                experience = Experience(observation=obs, next_observation=next_obs, action=action, reward=reward, done=done)  # do we really need to copy?
-                self.buffer.add(experience)
-                # end of step routine
-                obs = next_obs
-                step += 1
-                total_reward += reward
-                self.anneal_eps(step, n_steps)
-
-                if step % self.train_every_n_steps == 0:
-                    self.train()
-                    self.n_updates += 1
-                if step % self.target_update == 0:
-                    print('UPDATE')
-                    soft_update(self.q_net, self.target_q_net, tau=self.tau)
-
-            self.running_reward.append(total_reward)
-            if step % 10 == 0:
-                print(f'Step: {step} ({(step/n_steps)*100:.2f}%)\tRunning reward: {sum(list(self.running_reward))/len(self.running_reward):.2f}\t'
-                      f' eps: {self.eps:.4f}\tRunning loss: {sum(list(self.running_loss))/len(self.running_loss):.4f}\tUpdates:{self.n_updates}')
+    def on_step_end(self, n_steps):
+        self.anneal_eps(self.step, n_steps)
+        if self.step % self.target_update == 0:
+            print('UPDATE')
+            soft_update(self.q_net, self.target_q_net, tau=self.tau)
 
     def _training_routine(self, obs, next_obs, action):
         current_q_values = self.q_net(obs)
@@ -113,7 +90,7 @@ class QLearner(BaseLearner):
     def train(self):
         if len(self.buffer) < self.batch_size: return
         for _ in range(self.n_grad_steps):
-            experience = self.buffer.sample(self.batch_size, cer=self.train_every_n_steps)
+            experience = self.buffer.sample(self.batch_size, cer=self.train_every[-1])
             pred_q, target_q_raw = self._training_routine(experience.observation,
                                                           experience.next_observation,
                                                           experience.action)
@@ -127,8 +104,9 @@ if __name__ == '__main__':
     from environments.factory.simple_factory import SimpleFactory, DirtProperties, MovementProperties
     from algorithms.common import BaseDDQN
     from algorithms.vdn_learner import VDNLearner
+    from algorithms.udr_learner import UDRLearner
 
-    N_AGENTS = 2
+    N_AGENTS = 1
 
     dirt_props = DirtProperties(clean_amount=3, gain_amount=0.2, max_global_amount=30,
                                 max_local_amount=5, spawn_frequency=1, max_spawn_ratio=0.05)
@@ -138,7 +116,7 @@ if __name__ == '__main__':
     env = SimpleFactory(dirt_properties=dirt_props, movement_properties=move_props, n_agents=N_AGENTS, pomdp_radius=2,  max_steps=400, omit_agent_slice_in_obs=False, combin_agent_slices_in_obs=True)
 
     dqn, target_dqn = BaseDDQN(), BaseDDQN()
-    learner = VDNLearner(dqn, target_dqn, env, 40000, target_update=3500, lr=0.0007, gamma=0.99, n_agents=N_AGENTS, tau=0.95, max_grad_norm=10,
-                       train_every_n_steps=4, eps_end=0.025, n_grad_steps=1, reg_weight=0.1, exploration_fraction=0.25, batch_size=64)
+    learner = QLearner(dqn, target_dqn, env, 40000, target_update=3500, lr=0.0007, gamma=0.99, n_agents=N_AGENTS, tau=0.95, max_grad_norm=10,
+                       train_every=('step', 4), eps_end=0.025, n_grad_steps=1, reg_weight=0.1, exploration_fraction=0.25, batch_size=64)
     #learner.save(Path(__file__).parent / 'test' / 'testexperiment1337')
     learner.learn(100000)

algorithms/udr_learner.py

@@ -0,0 +1,178 @@
+import random
+from typing import Union, List
+from collections import deque
+import numpy as np
+import torch
+import torch.nn as nn
+from algorithms.common import BaseBuffer, Experience, BaseLearner, BaseDQN, mlp_maker
+from collections import defaultdict
+
+
+class UDRLBuffer(BaseBuffer):
+    def __init__(self, size):
+        super(UDRLBuffer, self).__init__(0)
+        self.experience = defaultdict(list)
+        self.size = size
+
+    def add(self, experience):
+        self.experience[experience.episode].append(experience)
+        if len(self.experience) > self.size:
+            self.sort_and_prune()
+
+    def select_time_steps(self, episode: List[Experience]):
+        T = len(episode)  # max horizon
+        t1 = random.randint(0, T - 1)
+        t2 = random.randint(t1 + 1, T)
+        return t1, t2, T
+
+    def sort_and_prune(self):
+        scores = []
+        for k, episode_experience in self.experience.items():
+            r = sum([e.reward for e in episode_experience])
+            scores.append((r, k))
+        sorted_scores = sorted(scores, reverse=True)
+        return sorted_scores
+
+    def sample(self, batch_size, cer=0):
+        random_episode_keys = random.choices(list(self.experience.keys()), k=batch_size)
+        lsts = (obs, desired_rewards, horizons, actions) = [], [], [], []
+        for ek in random_episode_keys:
+            episode = self.experience[ek]
+            t1, t2, T = self.select_time_steps(episode)
+            t2 = T  # TODO only good for episodic envs
+            observation = episode[t1].observation
+            desired_reward = sum([experience.reward for experience in episode[t1:t2]])
+            horizon = t2 - t1
+            action = episode[t1].action
+            for lst, val in zip(lsts, [observation, desired_reward, horizon, action]):
+                lst.append(val)
+        return (torch.stack([torch.from_numpy(o) for o in obs], 0).float(),
+                torch.tensor(desired_rewards).view(-1, 1).float(),
+                torch.tensor(horizons).view(-1, 1).float(),
+                torch.tensor(actions))
+
+
+class UDRLearner(BaseLearner):
+    # Upside Down Reinforcement Learner
+    def __init__(self, env, desired_reward, desired_horizon,
+                 behavior_fn=None, buffer_size=100, n_warm_up_episodes=8, best_x=20,
+                 batch_size=128, lr=1e-3, n_agents=1, train_every=('episode', 4), n_grad_steps=1):
+        super(UDRLearner, self).__init__(env, n_agents, train_every, n_grad_steps)
+        assert self.n_agents == 1, 'UDRL currently only supports single agent training'
+        self.behavior_fn = behavior_fn
+        self.buffer_size = buffer_size
+        self.n_warm_up_episodes = n_warm_up_episodes
+        self.buffer = UDRLBuffer(buffer_size)
+        self.batch_size = batch_size
+        self.mode = 'train'
+        self.best_x = best_x
+        self.desired_reward = desired_reward
+        self.desired_horizon = desired_horizon
+        self.lr = lr
+        self.optimizer = torch.optim.AdamW(self.behavior_fn.parameters(), lr=lr)
+
+        self.running_loss = deque(maxlen=self.n_grad_steps*5)
+
+    def sample_exploratory_commands(self):
+        top_x = self.buffer.sort_and_prune()[:self.best_x]
+        # The exploratory desired horizon dh0 is set to the mean of the lengths of the selected episodes
+        new_desired_horizon = np.mean([len(self.buffer.experience[k]) for _, k in top_x])
+        # save all top_X cumulative returns in a list
+        returns = [r for r, _ in top_x]
+        # from these returns calc the mean and std
+        mean_returns = np.mean([r for r, _ in top_x])
+        std_returns = np.std(returns)
+        # sample desired reward from a uniform distribution given the mean and the std
+        new_desired_reward = np.random.uniform(mean_returns, mean_returns + std_returns)
+        self.exploratory_commands = (new_desired_reward, new_desired_horizon)
+        return torch.tensor([[new_desired_reward]]).float(), torch.tensor([[new_desired_horizon]]).float()
+
+    def on_new_experience(self, experience):
+        self.buffer.add(experience)
+        self.desired_reward = self.desired_reward - torch.tensor(experience.reward).float().view(1, 1)
+
+    def on_step_end(self, n_steps):
+        one = torch.tensor([1.]).float().view(1, 1)
+        self.desired_horizon -= one
+        self.desired_horizon = self.desired_horizon if self.desired_horizon >= 1. else one
+
+    def on_episode_end(self, n_steps):
+        self.desired_reward, self.desired_horizon = self.sample_exploratory_commands()
+
+    def get_action(self, obs) -> Union[int, np.ndarray]:
+        o = torch.from_numpy(obs).unsqueeze(0) if self.n_agents <= 1 else torch.from_numpy(obs)
+        bf_out = self.behavior_fn(o.float(), self.desired_reward, self.desired_horizon)
+        dist = torch.distributions.Categorical(bf_out)
+        sample = dist.sample()
+        return [sample.item()]#[self.env.action_space.sample()]
+
+    def _backprop_loss(self, loss):
+        # log loss
+        self.running_loss.append(loss.item())
+        # Optimize the model
+        self.optimizer.zero_grad()
+        loss.backward()
+        #torch.nn.utils.clip_grad_norm_(self.behavior_fn.parameters(), 10)
+        self.optimizer.step()
+
+    def train(self):
+        if len(self.buffer) < self.n_warm_up_episodes: return
+        for _ in range(self.n_grad_steps):
+            experience = self.buffer.sample(self.batch_size)
+            bf_out = self.behavior_fn(*experience[:3])
+            labels = experience[-1]
+            #print(labels.shape)
+            loss = nn.CrossEntropyLoss()(bf_out, labels.squeeze())
+            mean_entropy = torch.distributions.Categorical(bf_out).entropy().mean()
+            self._backprop_loss(loss - 0.03*mean_entropy)
+        print(f'Running loss: {np.mean(list(self.running_loss)):.3f}\tRunning reward: {np.mean(self.running_reward):.2f}'
+              f'\td_r: {self.desired_reward.item():.2f}\ttd_h: {self.desired_horizon.item()}')
+
+
+class BF(BaseDQN):
+    def __init__(self, dims=[5*5*3, 64]):
+        super(BF, self).__init__(dims)
+        self.net = mlp_maker(dims, activation_last='identity')
+        self.command_net = mlp_maker([2, 64], activation_last='sigmoid')
+        self.common_branch = mlp_maker([64, 64, 64, 9])
+
+
+    def forward(self, observation, desired_reward, horizon):
+        command = torch.cat((desired_reward*(0.02), horizon*(0.01)), dim=-1)
+        obs_out = self.net(torch.flatten(observation, start_dim=1))
+        command_out = self.command_net(command)
+        combined = obs_out*command_out
+        out = self.common_branch(combined)
+        return torch.softmax(out, -1)
+
+
+if __name__ == '__main__':
+    from environments.factory.simple_factory import SimpleFactory, DirtProperties, MovementProperties
+    from algorithms.common import BaseDDQN
+    from algorithms.vdn_learner import VDNLearner
+
+    N_AGENTS = 1
+
+    dirt_props = DirtProperties(clean_amount=3, gain_amount=0.2, max_global_amount=30,
+                                max_local_amount=5, spawn_frequency=1, max_spawn_ratio=0.05)
+    move_props = MovementProperties(allow_diagonal_movement=True,
+                                    allow_square_movement=True,
+                                    allow_no_op=False)
+    env = SimpleFactory(dirt_properties=dirt_props, movement_properties=move_props, n_agents=N_AGENTS, pomdp_radius=2,
+                        max_steps=400, omit_agent_slice_in_obs=False, combin_agent_slices_in_obs=True)
+
+    bf = BF()
+    desired_reward = torch.tensor([200.]).view(1, 1).float()
+    desired_horizon = torch.tensor([400.]).view(1, 1).float()
+    learner = UDRLearner(env, behavior_fn=bf,
+                         train_every=('episode', 2),
+                         buffer_size=40,
+                         best_x=10,
+                         lr=1e-3,
+                         batch_size=64,
+                         n_warm_up_episodes=12,
+                         n_grad_steps=4,
+                         desired_reward=desired_reward,
+                         desired_horizon=desired_horizon)
+    #learner.save(Path(__file__).parent / 'test' / 'testexperiment1337')
+    learner.learn(500000)