agent.py 49.7 KB
Newer Older
1 2 3 4 5 6 7
import mxnet as mx
import numpy as np
import time
import os
import sys
import util
import matplotlib.pyplot as plt
Nicola Gatto's avatar
Nicola Gatto committed
8 9
import pyprind
from cnnarch_logger import ArchLogger
10
from replay_memory import ReplayMemoryBuilder
Nicola Gatto's avatar
Nicola Gatto committed
11 12 13 14
from strategy import StrategyBuilder
from util import copy_net, get_loss_function,\
    copy_net_with_two_inputs, DdpgTrainingStats, DqnTrainingStats,\
    make_directory_if_not_exist
15 16
from mxnet import nd, gluon, autograd

Nicola Gatto's avatar
Nicola Gatto committed
17 18 19 20

class Agent(object):
    def __init__(
        self,
21 22
        environment,
        replay_memory_params,
Nicola Gatto's avatar
Nicola Gatto committed
23
        strategy_params,
24
        state_dim,
Nicola Gatto's avatar
Nicola Gatto committed
25
        action_dim,
26 27 28 29
        ctx=None,
        discount_factor=.9,
        training_episodes=50,
        train_interval=1,
Nicola Gatto's avatar
Nicola Gatto committed
30
        start_training=0,
31
        snapshot_interval=200,
Nicola Gatto's avatar
Nicola Gatto committed
32
        agent_name='Agent',
33
        max_episode_step=99999,
Nicola Gatto's avatar
Nicola Gatto committed
34
        evaluation_samples=1000,
35 36
        output_directory='model_parameters',
        verbose=True,
Nicola Gatto's avatar
Nicola Gatto committed
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
        target_score=None
    ):
        assert 0 < discount_factor <= 1,\
            'Discount factor must be between 0 and 1'
        assert train_interval > 0, 'Train interval must be greater 0'
        assert snapshot_interval > 0, 'Snapshot interval must be greater 0'
        assert max_episode_step > 0,\
            'Maximal steps per episode must be greater 0'
        assert training_episodes > 0, 'Trainings episode must be greater 0'
        assert replay_memory_params is not None,\
            'Replay memory parameter not set'
        assert type(state_dim) is tuple, 'State dimension is not a tuple'
        assert type(action_dim) is tuple, 'Action dimension is not a tuple'

        self._logger = ArchLogger.get_logger()
        self._ctx = mx.gpu() if ctx == 'gpu' else mx.cpu()
        self._environment = environment
        self._discount_factor = discount_factor
        self._training_episodes = training_episodes
        self._train_interval = train_interval
        self._verbose = verbose
        self._state_dim = state_dim
59 60

        replay_memory_params['state_dim'] = state_dim
Nicola Gatto's avatar
Nicola Gatto committed
61 62
        replay_memory_params['action_dim'] = action_dim
        self._replay_memory_params = replay_memory_params
63
        rm_builder = ReplayMemoryBuilder()
Nicola Gatto's avatar
Nicola Gatto committed
64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
        self._memory = rm_builder.build_by_params(**replay_memory_params)
        self._minibatch_size = self._memory.sample_size
        self._action_dim = action_dim

        strategy_params['action_dim'] = self._action_dim
        self._strategy_params = strategy_params
        strategy_builder = StrategyBuilder()
        self._strategy = strategy_builder.build_by_params(**strategy_params)
        self._agent_name = agent_name
        self._snapshot_interval = snapshot_interval
        self._creation_time = time.time()
        self._max_episode_step = max_episode_step
        self._start_training = start_training
        self._output_directory = output_directory
        self._target_score = target_score

        self._evaluation_samples = evaluation_samples
        self._best_avg_score = -np.infty
        self._best_net = None

        self._interrupt_flag = False
        self._training_stats = None
86 87

        # Training Context
Nicola Gatto's avatar
Nicola Gatto committed
88 89
        self._current_episode = 0
        self._total_steps = 0
90

Nicola Gatto's avatar
Nicola Gatto committed
91 92 93
    @property
    def current_episode(self):
        return self._current_episode
94

Nicola Gatto's avatar
Nicola Gatto committed
95 96 97
    @property
    def environment(self):
        return self._environment
98

Nicola Gatto's avatar
Nicola Gatto committed
99
    def save_config_file(self):
100
        import json
Nicola Gatto's avatar
Nicola Gatto committed
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116
        make_directory_if_not_exist(self._output_directory)
        filename = os.path.join(self._output_directory, 'config.json')
        config = self._make_config_dict()
        with open(filename, mode='w') as fp:
            json.dump(config, fp, indent=4)

    def set_interrupt_flag(self, interrupt):
        self._interrupt_flag = interrupt

    def _interrupt_training(self):
        import pickle
        self._logger.info('Training interrupted; Store state for resuming')
        session_dir = self._get_session_dir()
        agent_session_file = os.path.join(session_dir, 'agent.p')
        logger = self._logger

117 118
        self._training_stats.save_stats(self._output_directory, episode=self._current_episode)

Nicola Gatto's avatar
Nicola Gatto committed
119 120 121 122 123 124 125 126 127 128
        self._make_pickle_ready(session_dir)

        with open(agent_session_file, 'wb') as f:
            pickle.dump(self, f, protocol=2)
        logger.info('State successfully stored')

    def _make_pickle_ready(self, session_dir):
        del self._training_stats.logger
        self._environment.close()
        self._environment = None
129 130
        self._export_net(self._best_net, 'best_net', filedir=session_dir)
        self._logger = None
Nicola Gatto's avatar
Nicola Gatto committed
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
        self._best_net = None

    def _make_config_dict(self):
        config = dict()
        config['state_dim'] = self._state_dim
        config['action_dim'] = self._action_dim
        config['ctx'] = str(self._ctx)
        config['discount_factor'] = self._discount_factor
        config['strategy_params'] = self._strategy_params
        config['replay_memory_params'] = self._replay_memory_params
        config['training_episodes'] = self._training_episodes
        config['start_training'] = self._start_training
        config['evaluation_samples'] = self._evaluation_samples
        config['train_interval'] = self._train_interval
        config['snapshot_interval'] = self._snapshot_interval
        config['agent_name'] = self._agent_name
        config['max_episode_step'] = self._max_episode_step
        config['output_directory'] = self._output_directory
        config['verbose'] = self._verbose
        config['target_score'] = self._target_score
        return config

    def _adjust_optimizer_params(self, optimizer_params):
        if 'weight_decay' in optimizer_params:
            optimizer_params['wd'] = optimizer_params['weight_decay']
            del optimizer_params['weight_decay']
        if 'learning_rate_decay' in optimizer_params:
            min_learning_rate = 1e-8
            if 'learning_rate_minimum' in optimizer_params:
                min_learning_rate = optimizer_params['learning_rate_minimum']
                del optimizer_params['learning_rate_minimum']
            optimizer_params['lr_scheduler'] = mx.lr_scheduler.FactorScheduler(
                optimizer_params['step_size'],
                factor=optimizer_params['learning_rate_decay'],
                stop_factor_lr=min_learning_rate)
            del optimizer_params['step_size']
            del optimizer_params['learning_rate_decay']

        return optimizer_params

    def _sample_from_memory(self):
        states, actions, rewards, next_states, terminals\
            = self._memory.sample(batch_size=self._minibatch_size)
        states = nd.array(states, ctx=self._ctx)
        actions = nd.array(actions, ctx=self._ctx)
        rewards = nd.array(rewards, ctx=self._ctx)
        next_states = nd.array(next_states, ctx=self._ctx)
        terminals = nd.array(terminals, ctx=self._ctx)
        return states, actions, rewards, next_states, terminals

    def evaluate(self, target=None, sample_games=100, verbose=True):
182 183 184
        if sample_games <= 0:
            return 0

Nicola Gatto's avatar
Nicola Gatto committed
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
        target = self._target_score if target is None else target
        if target:
            target_achieved = 0
        total_reward = 0

        self._logger.info('Sampling from {} games...'.format(sample_games))
        for g in pyprind.prog_bar(range(sample_games)):
            state = self._environment.reset()
            step = 0
            game_reward = 0
            terminal = False
            while not terminal and (step < self._max_episode_step):
                action = self.get_next_action(state)
                state, reward, terminal, _ = self._environment.step(action)
                game_reward += reward
                step += 1

            if verbose:
                info = 'Game %d: Reward %f' % (g, game_reward)
                self._logger.debug(info)
            if target:
                if game_reward >= target:
                    target_achieved += 1
            total_reward += game_reward

        avg_reward = float(total_reward)/float(sample_games)
        info = 'Avg. Reward: %f' % avg_reward
        if target:
            target_achieved_ratio = int(
                (float(target_achieved)/float(sample_games))*100)
            info += '; Target Achieved in %d%% of games'\
                % (target_achieved_ratio)

        if verbose:
            self._logger.info(info)
        return avg_reward

    def _do_snapshot_if_in_interval(self, episode):
        do_snapshot =\
            (episode != 0 and (episode % self._snapshot_interval == 0))
        if do_snapshot:
            self.save_parameters(episode=episode)
            self._evaluate()

    def _evaluate(self, verbose=True):
        avg_reward = self.evaluate(
            sample_games=self._evaluation_samples, verbose=False)
        info = 'Evaluation -> Average Reward in {} games: {}'.format(
            self._evaluation_samples, avg_reward)

        if self._best_avg_score is None or self._best_avg_score <= avg_reward:
            self._save_current_as_best_net()
            self._best_avg_score = avg_reward
        if verbose:
            self._logger.info(info)

    def _is_target_reached(self, avg_reward):
        return self._target_score is not None\
            and avg_reward > self._target_score

    def _do_training(self):
        return (self._total_steps % self._train_interval == 0) and\
            (self._memory.is_sample_possible(self._minibatch_size)) and\
            (self._current_episode >= self._start_training)

    def _check_interrupt_routine(self):
        if self._interrupt_flag:
            self._interrupt_flag = False
            self._interrupt_training()
            return True
        return False

    def _is_target_reached(self, avg_reward):
        return self._target_score is not None\
            and avg_reward > self._target_score

261
    def _export_net(self, net, filename, filedir=None, episode=None):
Nicola Gatto's avatar
Nicola Gatto committed
262 263 264
        assert self._output_directory
        assert isinstance(net, gluon.HybridBlock)
        make_directory_if_not_exist(self._output_directory)
265 266
        filedir = self._output_directory if filedir is None else filedir
        filename = os.path.join(filedir, filename)
Nicola Gatto's avatar
Nicola Gatto committed
267

268
        if episode is not None:
Nicola Gatto's avatar
Nicola Gatto committed
269 270
            filename = filename + '-ep{}'.format(episode)

271
        net.export(filename, epoch=0)
272
        net.save_parameters(filename + '.params')
Nicola Gatto's avatar
Nicola Gatto committed
273

274 275 276
    def export_best_network(self, path=None, epoch=0):
        path = os.path.join(self._output_directory, 'best_network')\
            if path is None else path
Nicola Gatto's avatar
Nicola Gatto committed
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371
        self._logger.info(
            'Saving best network with average reward of {}'.format(
                self._best_avg_score))
        self._best_net.export(path, epoch=epoch)

    def _get_session_dir(self):
        session_dir = os.path.join(
            self._output_directory, '.interrupted_session')
        make_directory_if_not_exist(session_dir)
        return session_dir

    def _save_current_as_best_net(self):
        raise NotImplementedError

    def get_next_action(self, state):
        raise NotImplementedError

    def save_parameters(self, episode):
        raise NotImplementedError

    def train(self, episodes=None):
        raise NotImplementedError


class DdpgAgent(Agent):
    def __init__(
        self,
        actor,
        critic,
        environment,
        replay_memory_params,
        strategy_params,
        state_dim,
        action_dim,
        soft_target_update_rate=.001,
        actor_optimizer='adam',
        actor_optimizer_params={'learning_rate': 0.0001},
        critic_optimizer='adam',
        critic_optimizer_params={'learning_rate': 0.001},
        ctx=None,
        discount_factor=.9,
        training_episodes=50,
        start_training=20,
        train_interval=1,
        snapshot_interval=200,
        agent_name='DdpgAgent',
        max_episode_step=9999,
        evaluation_samples=100,
        output_directory='model_parameters',
        verbose=True,
        target_score=None
    ):
        super(DdpgAgent, self).__init__(
            environment=environment, replay_memory_params=replay_memory_params,
            strategy_params=strategy_params, state_dim=state_dim,
            action_dim=action_dim, ctx=ctx, discount_factor=discount_factor,
            training_episodes=training_episodes, start_training=start_training,
            train_interval=train_interval,
            snapshot_interval=snapshot_interval, agent_name=agent_name,
            max_episode_step=max_episode_step,
            output_directory=output_directory, verbose=verbose,
            target_score=target_score, evaluation_samples=evaluation_samples)
        assert critic is not None, 'Critic not set'
        assert actor is not None, 'Actor is not set'
        assert soft_target_update_rate > 0,\
            'Target update must be greater zero'
        assert actor_optimizer is not None, 'No actor optimizer set'
        assert critic_optimizer is not None, 'No critic optimizer set'

        self._actor = actor
        self._critic = critic

        self._actor_target = self._copy_actor()
        self._critic_target = self._copy_critic()

        self._actor_optimizer = actor_optimizer
        self._actor_optimizer_params = self._adjust_optimizer_params(
            actor_optimizer_params)

        self._critic_optimizer = critic_optimizer
        self._critic_optimizer_params = self._adjust_optimizer_params(
            critic_optimizer_params)

        self._soft_target_update_rate = soft_target_update_rate

        self._logger.info(
            'Agent created with following parameters: {}'.format(
                self._make_config_dict()))

        self._best_net = self._copy_actor()

        self._training_stats = DdpgTrainingStats(self._training_episodes)

    def _make_pickle_ready(self, session_dir):
        super(DdpgAgent, self)._make_pickle_ready(session_dir)
372
        self._export_net(self._actor, 'current_actor')
373

374
        self._export_net(self._actor, 'actor', filedir=session_dir)
Nicola Gatto's avatar
Nicola Gatto committed
375
        self._actor = None
376
        self._export_net(self._critic, 'critic', filedir=session_dir)
Nicola Gatto's avatar
Nicola Gatto committed
377
        self._critic = None
378 379
        self._export_net(
            self._actor_target, 'actor_target', filedir=session_dir)
Nicola Gatto's avatar
Nicola Gatto committed
380
        self._actor_target = None
381 382
        self._export_net(
            self._critic_target, 'critic_target', filedir=session_dir)
Nicola Gatto's avatar
Nicola Gatto committed
383
        self._critic_target = None
384 385

    @classmethod
Nicola Gatto's avatar
Nicola Gatto committed
386
    def resume_from_session(cls, session_dir, actor, critic, environment):
387 388 389 390 391 392 393
        import pickle
        if not os.path.exists(session_dir):
            raise ValueError('Session directory does not exist')

        files = dict()
        files['agent'] = os.path.join(session_dir, 'agent.p')
        files['best_net_params'] = os.path.join(session_dir, 'best_net.params')
Nicola Gatto's avatar
Nicola Gatto committed
394 395 396 397 398 399
        files['actor_net_params'] = os.path.join(session_dir, 'actor.params')
        files['actor_target_net_params'] = os.path.join(
            session_dir, 'actor_target.params')
        files['critic_net_params'] = os.path.join(session_dir, 'critic.params')
        files['critic_target_net_params'] = os.path.join(
            session_dir, 'critic_target.params')
400 401 402

        for file in files.values():
            if not os.path.exists(file):
Nicola Gatto's avatar
Nicola Gatto committed
403 404 405
                raise ValueError(
                    'Session directory is not complete: {} is missing'
                    .format(file))
406 407 408 409

        with open(files['agent'], 'rb') as f:
            agent = pickle.load(f)

Nicola Gatto's avatar
Nicola Gatto committed
410 411 412 413 414 415 416 417 418 419 420 421 422 423
        agent._environment = environment

        agent._actor = actor
        agent._actor.load_parameters(files['actor_net_params'], agent._ctx)
        agent._actor.hybridize()
        agent._actor(nd.random_normal(
            shape=((1,) + agent._state_dim), ctx=agent._ctx))

        agent._best_net = copy_net(agent._actor, agent._state_dim, agent._ctx)
        agent._best_net.load_parameters(files['best_net_params'], agent._ctx)

        agent._actor_target = copy_net(
            agent._actor, agent._state_dim, agent._ctx)
        agent._actor_target.load_parameters(files['actor_target_net_params'])
424

Nicola Gatto's avatar
Nicola Gatto committed
425 426 427 428 429 430 431 432 433 434 435 436 437 438
        agent._critic = critic
        agent._critic.load_parameters(files['critic_net_params'], agent._ctx)
        agent._critic.hybridize()
        agent._critic(
            nd.random_normal(shape=((1,) + agent._state_dim), ctx=agent._ctx),
            nd.random_normal(shape=((1,) + agent._action_dim), ctx=agent._ctx))

        agent._critic_target = copy_net_with_two_inputs(
            agent._critic, agent._state_dim, agent._action_dim, agent._ctx)
        agent._critic_target.load_parameters(files['critic_target_net_params'])

        agent._logger = ArchLogger.get_logger()
        agent._training_stats.logger = ArchLogger.get_logger()
        agent._logger.info('Agent was retrieved; Training can be continued')
439 440 441

        return agent

Nicola Gatto's avatar
Nicola Gatto committed
442 443
    def _save_current_as_best_net(self):
        self._best_net = self._copy_actor()
444

Nicola Gatto's avatar
Nicola Gatto committed
445 446 447
    def get_next_action(self, state):
        action = self._actor(nd.array([state], ctx=self._ctx))
        return action[0].asnumpy()
448

Nicola Gatto's avatar
Nicola Gatto committed
449
    def save_parameters(self, episode):
450 451
        self._export_net(
            self._actor, self._agent_name + '_actor', episode=episode)
452

Nicola Gatto's avatar
Nicola Gatto committed
453 454 455 456 457
    def train(self, episodes=None):
        self.save_config_file()
        self._logger.info("--- Start DDPG training ---")
        episodes = \
            episodes if episodes is not None else self._training_episodes
458

Nicola Gatto's avatar
Nicola Gatto committed
459 460 461 462 463 464 465
        resume = (self._current_episode > 0)
        if resume:
            self._logger.info("Training session resumed")
            self._logger.info(
                "Starting from episode {}".format(self._current_episode))
        else:
            self._training_stats = DdpgTrainingStats(episodes)
466

467 468 469
            # Initialize target Q' and mu'
            self._actor_target = self._copy_actor()
            self._critic_target = self._copy_critic()
470

Nicola Gatto's avatar
Nicola Gatto committed
471 472
        # Initialize l2 loss for critic network
        l2_loss = gluon.loss.L2Loss()
473

Nicola Gatto's avatar
Nicola Gatto committed
474 475 476 477 478 479 480
        # Initialize critic and actor trainer
        trainer_actor = gluon.Trainer(
            self._actor.collect_params(), self._actor_optimizer,
            self._actor_optimizer_params)
        trainer_critic = gluon.Trainer(
            self._critic.collect_params(), self._critic_optimizer,
            self._critic_optimizer_params)
481

Nicola Gatto's avatar
Nicola Gatto committed
482 483 484 485 486
        # For episode=1..n
        while self._current_episode < episodes:
            # Check interrupt flag
            if self._check_interrupt_routine():
                return False
487

Nicola Gatto's avatar
Nicola Gatto committed
488 489 490 491 492 493 494 495
            # Initialize new episode
            step = 0
            episode_reward = 0
            start = time.time()
            episode_critic_loss = 0
            episode_actor_loss = 0
            episode_avg_q_value = 0
            training_steps = 0
496

Nicola Gatto's avatar
Nicola Gatto committed
497 498 499 500 501 502 503 504 505
            # Get initialial observation state s
            state = self._environment.reset()

            # For step=1..T
            while step < self._max_episode_step:
                # Select an action a = mu(s) + N(step) according to current
                #  actor and exploration noise N according to strategy
                action = self._strategy.select_action(
                    self.get_next_action(state))
506
                self._strategy.decay(self._current_episode)
Nicola Gatto's avatar
Nicola Gatto committed
507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549

                # Execute action a and observe reward r and next state ns
                next_state, reward, terminal, _ = \
                    self._environment.step(action)

                self._logger.debug(
                    'Applied action {} with reward {}'.format(action, reward))

                # Store transition (s,a,r,ns) in replay buffer
                self._memory.append(
                    state, action, reward, next_state, terminal)

                if self._do_training():
                    # Sample random minibatch of b transitions
                    # (s_i, a_i, r_i, s_(i+1)) from replay buffer
                    states, actions, rewards, next_states, terminals =\
                         self._sample_from_memory()

                    actor_target_actions = self._actor_target(next_states)
                    critic_target_qvalues = self._critic_target(
                        next_states, actor_target_actions)

                    rewards = rewards.reshape(self._minibatch_size, 1)
                    terminals = terminals.reshape(self._minibatch_size, 1)

                    # y = r_i + discount * Q'(s_(i+1), mu'(s_(i+1)))
                    y = rewards + (1.0 - terminals) * self._discount_factor\
                        * critic_target_qvalues

                    # Train the critic network
                    with autograd.record():
                        qvalues = self._critic(states, actions)
                        critic_loss = l2_loss(qvalues, y)
                    critic_loss.backward()
                    trainer_critic.step(self._minibatch_size)

                    # Train the actor network
                    # Temporary critic so that gluon trainer does not mess
                    # with critic parameters
                    tmp_critic = self._copy_critic()
                    with autograd.record():
                        # For maximizing qvalues we have to multiply with -1
                        # as we use a minimizer
550 551
                        actor_loss = -tmp_critic(
                            states, self._actor(states)).mean()
Nicola Gatto's avatar
Nicola Gatto committed
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568
                    actor_loss.backward()
                    trainer_actor.step(self._minibatch_size)

                    # Update target networks:
                    self._actor_target = self._soft_update(
                        self._actor, self._actor_target,
                        self._soft_target_update_rate)
                    self._critic_target = self._soft_update(
                        self._critic, self._critic_target,
                        self._soft_target_update_rate)

                    # Update statistics
                    episode_critic_loss +=\
                        np.sum(critic_loss.asnumpy()) / self._minibatch_size
                    episode_actor_loss +=\
                        np.sum(actor_loss.asnumpy()) / self._minibatch_size
                    episode_avg_q_value +=\
569
                        np.sum(qvalues.asnumpy()) / self._minibatch_size
570

Nicola Gatto's avatar
Nicola Gatto committed
571
                    training_steps += 1
572

Nicola Gatto's avatar
Nicola Gatto committed
573 574 575 576
                episode_reward += reward
                step += 1
                self._total_steps += 1
                state = next_state
577

Nicola Gatto's avatar
Nicola Gatto committed
578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606
                if terminal:
                    # Reset the strategy
                    self._strategy.reset()
                    break

            # Log the episode results
            episode_actor_loss = 0 if training_steps == 0\
                else (episode_actor_loss / training_steps)
            episode_critic_loss = 0 if training_steps == 0\
                else (episode_critic_loss / training_steps)
            episode_avg_q_value = 0 if training_steps == 0\
                else (episode_avg_q_value / training_steps)

            avg_reward = self._training_stats.log_episode(
                self._current_episode, start, training_steps,
                episode_actor_loss, episode_critic_loss, episode_avg_q_value,
                self._strategy.cur_eps, episode_reward)

            self._do_snapshot_if_in_interval(self._current_episode)

            if self._is_target_reached(avg_reward):
                self._logger.info(
                    'Target score is reached in average; Training is stopped')
                break

            self._current_episode += 1

        self._evaluate()
        self.save_parameters(episode=self._current_episode)
607
        self.export_best_network()
Nicola Gatto's avatar
Nicola Gatto committed
608 609 610 611 612 613 614 615 616 617 618 619
        self._training_stats.save_stats(self._output_directory)
        self._logger.info('--------- Training finished ---------')
        return True

    def _make_config_dict(self):
        config = super(DdpgAgent, self)._make_config_dict()
        config['soft_target_update_rate'] = self._soft_target_update_rate
        config['actor_optimizer'] = self._actor_optimizer
        config['actor_optimizer_params'] = self._actor_optimizer_params
        config['critic_optimizer'] = self._critic_optimizer
        config['critic_optimizer_params'] = self._critic_optimizer_params
        return config
620

Nicola Gatto's avatar
Nicola Gatto committed
621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642
    def _soft_update(self, net, target, tau):
        net_params = [p.data() for _, p in net.collect_params().items()]
        for i, (_, p) in enumerate(target.collect_params().items()):
            target_params = p.data()
            p.set_data((1.0 - tau) * target_params + tau * net_params[i])
        return target

    def _copy_actor(self):
        assert self._actor is not None
        assert self._ctx is not None
        assert type(self._state_dim) is tuple
        return copy_net(self._actor, self._state_dim, ctx=self._ctx)

    def _copy_critic(self):
        assert self._critic is not None
        assert self._ctx is not None
        assert type(self._state_dim) is tuple
        assert type(self._action_dim) is tuple
        return copy_net_with_two_inputs(
            self._critic, self._state_dim, self._action_dim, ctx=self._ctx)


643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995
class TwinDelayedDdpgAgent(DdpgAgent):
    def __init__(
        self,
        actor,
        critic,
        environment,
        replay_memory_params,
        strategy_params,
        state_dim,
        action_dim,
        soft_target_update_rate=.001,
        actor_optimizer='adam',
        actor_optimizer_params={'learning_rate': 0.0001},
        critic_optimizer='adam',
        critic_optimizer_params={'learning_rate': 0.001},
        ctx=None,
        discount_factor=.9,
        training_episodes=50,
        start_training=20,
        train_interval=1,
        snapshot_interval=200,
        agent_name='DdpgAgent',
        max_episode_step=9999,
        evaluation_samples=100,
        output_directory='model_parameters',
        verbose=True,
        target_score=None,
        policy_noise=0.2,
        noise_clip=0.5,
        policy_delay=2
    ):
        super(TwinDelayedDdpgAgent, self).__init__(
            environment=environment, replay_memory_params=replay_memory_params,
            strategy_params=strategy_params, state_dim=state_dim,
            action_dim=action_dim, ctx=ctx, discount_factor=discount_factor,
            training_episodes=training_episodes, start_training=start_training,
            train_interval=train_interval,
            snapshot_interval=snapshot_interval, agent_name=agent_name,
            max_episode_step=max_episode_step,
            output_directory=output_directory, verbose=verbose,
            target_score=target_score, evaluation_samples=evaluation_samples,
            critic=critic, soft_target_update_rate=soft_target_update_rate,
            actor=actor, actor_optimizer=actor_optimizer,
            actor_optimizer_params=actor_optimizer_params,
            critic_optimizer=critic_optimizer,
            critic_optimizer_params=critic_optimizer_params)

        self._policy_noise = policy_noise
        self._noise_clip = noise_clip
        self._policy_delay = policy_delay

        self._critic2 = self._critic.__class__()
        self._critic2.collect_params().initialize(
            mx.init.Normal(), ctx=self._ctx)
        self._critic2.hybridize()
        self._critic2(nd.ones((1,) + state_dim, ctx=self._ctx),
                      nd.ones((1,) + action_dim, ctx=self._ctx))

        self._critic2_target = self._copy_critic2()

        self._critic2_optimizer = critic_optimizer
        self._critic2_optimizer_params = self._adjust_optimizer_params(
            critic_optimizer_params)

    def _make_pickle_ready(self, session_dir):
        super(TwinDelayedDdpgAgent, self)._make_pickle_ready(session_dir)
        self._export_net(self._critic2, 'critic2', filedir=session_dir)
        self._critic2 = None
        self._export_net(
            self._critic2_target, 'critic2_target', filedir=session_dir)
        self._critic2_target = None

    @classmethod
    def resume_from_session(cls, session_dir, actor, critic, environment):
        import pickle
        if not os.path.exists(session_dir):
            raise ValueError('Session directory does not exist')

        files = dict()
        files['agent'] = os.path.join(session_dir, 'agent.p')
        files['best_net_params'] = os.path.join(session_dir, 'best_net.params')
        files['actor_net_params'] = os.path.join(session_dir, 'actor.params')
        files['actor_target_net_params'] = os.path.join(
            session_dir, 'actor_target.params')
        files['critic_net_params'] = os.path.join(session_dir, 'critic.params')
        files['critic_target_net_params'] = os.path.join(
            session_dir, 'critic_target.params')
        files['critic2_net_params'] = os.path.join(
            session_dir, 'critic2.params')
        files['critic2_target_net_params'] = os.path.join(
            session_dir, 'critic2_target.params')

        for file in files.values():
            if not os.path.exists(file):
                raise ValueError(
                    'Session directory is not complete: {} is missing'
                    .format(file))

        with open(files['agent'], 'rb') as f:
            agent = pickle.load(f)

        agent._environment = environment

        agent._actor = actor
        agent._actor.load_parameters(files['actor_net_params'], agent._ctx)
        agent._actor.hybridize()
        agent._actor(nd.random_normal(
            shape=((1,) + agent._state_dim), ctx=agent._ctx))

        agent._best_net = copy_net(agent._actor, agent._state_dim, agent._ctx)
        agent._best_net.load_parameters(files['best_net_params'], agent._ctx)

        agent._actor_target = copy_net(
            agent._actor, agent._state_dim, agent._ctx)
        agent._actor_target.load_parameters(files['actor_target_net_params'])

        agent._critic = critic
        agent._critic.load_parameters(files['critic_net_params'], agent._ctx)
        agent._critic.hybridize()
        agent._critic(
            nd.random_normal(shape=((1,) + agent._state_dim), ctx=agent._ctx),
            nd.random_normal(shape=((1,) + agent._action_dim), ctx=agent._ctx))

        agent._critic_target = copy_net_with_two_inputs(
            agent._critic, agent._state_dim, agent._action_dim, agent._ctx)
        agent._critic_target.load_parameters(files['critic_target_net_params'])

        agent._critic2 = copy_net_with_two_inputs(
            agent._critic, agent._state_dim, agent._action_dim, agent._ctx)
        agent._critic2.load_parameters(files['critic2_net_params'], agent._ctx)
        agent._critic2.hybridize()
        agent._critic2(
            nd.random_normal(shape=((1,) + agent._state_dim), ctx=agent._ctx),
            nd.random_normal(shape=((1,) + agent._action_dim), ctx=agent._ctx))

        agent._critic2_target = copy_net_with_two_inputs(
            agent._critic2, agent._state_dim, agent._action_dim, agent._ctx)
        agent._critic2_target.load_parameters(
            files['critic2_target_net_params'])

        agent._logger = ArchLogger.get_logger()
        agent._training_stats.logger = ArchLogger.get_logger()
        agent._logger.info('Agent was retrieved; Training can be continued')

        return agent

    def _copy_critic2(self):
        assert self._critic2 is not None
        assert self._ctx is not None
        assert type(self._state_dim) is tuple
        assert type(self._action_dim) is tuple
        return copy_net_with_two_inputs(
            self._critic2, self._state_dim, self._action_dim, ctx=self._ctx)

    def train(self, episodes=None):
        self.save_config_file()
        self._logger.info("--- Start TwinDelayedDDPG training ---")
        episodes = \
            episodes if episodes is not None else self._training_episodes

        resume = (self._current_episode > 0)
        if resume:
            self._logger.info("Training session resumed")
            self._logger.info(
                "Starting from episode {}".format(self._current_episode))
        else:
            self._training_stats = DdpgTrainingStats(episodes)

            # Initialize target Q1' and Q2' and mu'
            self._actor_target = self._copy_actor()
            self._critic_target = self._copy_critic()
            self._critic2_target = self._copy_critic2()

        # Initialize l2 loss for critic network
        l2_loss = gluon.loss.L2Loss()

        # Initialize critic and actor trainer
        trainer_actor = gluon.Trainer(
            self._actor.collect_params(), self._actor_optimizer,
            self._actor_optimizer_params)
        trainer_critic = gluon.Trainer(
            self._critic.collect_params(), self._critic_optimizer,
            self._critic_optimizer_params)
        trainer_critic2 = gluon.Trainer(
            self._critic2.collect_params(), self._critic2_optimizer,
            self._critic2_optimizer_params)

        # For episode=1..n
        while self._current_episode < episodes:
            # Check interrupt flag
            if self._check_interrupt_routine():
                return False

            # Initialize new episode
            step = 0
            episode_reward = 0
            start = time.time()
            episode_critic_loss = 0
            episode_actor_loss = 0
            episode_avg_q_value = 0
            training_steps = 0
            actor_updates = 0

            # Get initialial observation state s
            state = self._environment.reset()

            # For step=1..T
            while step < self._max_episode_step:
                # Select an action a = mu(s) + N(step) according to current
                #  actor and exploration noise N according to strategy
                action = self._strategy.select_action(
                    self.get_next_action(state))
                self._strategy.decay(self._current_episode)

                # Execute action a and observe reward r and next state ns
                next_state, reward, terminal, _ = \
                    self._environment.step(action)

                self._logger.debug(
                    'Applied action {} with reward {}'.format(action, reward))

                # Store transition (s,a,r,ns) in replay buffer
                self._memory.append(
                    state, action, reward, next_state, terminal)

                if self._do_training():
                    # Sample random minibatch of b transitions
                    # (s_i, a_i, r_i, s_(i+1)) from replay buffer
                    states, actions, rewards, next_states, terminals =\
                         self._sample_from_memory()

                    clipped_noise = nd.array(
                        np.clip(
                            np.random.normal(
                                loc=0, scale=self._policy_noise,
                                size=self._minibatch_size
                            ).reshape(self._minibatch_size, 1),
                            -self._noise_clip,
                            self._noise_clip
                        ),
                        ctx=self._ctx
                    )
                    target_action = np.clip(
                        self._actor_target(next_states) + clipped_noise,
                        self._strategy._action_low,
                        self._strategy._action_high)

                    rewards = rewards.reshape(self._minibatch_size, 1)
                    terminals = terminals.reshape(self._minibatch_size, 1)

                    target_qvalues1 = self._critic_target(next_states,
                                                          target_action)
                    target_qvalues2 = self._critic2_target(next_states,
                                                           target_action)
                    target_qvalues = nd.minimum(target_qvalues1,
                                                target_qvalues2)
                    y = rewards + (1 - terminals) * self._discount_factor\
                        * target_qvalues

                    with autograd.record():
                        qvalues1 = self._critic(states, actions)
                        critic1_loss = l2_loss(qvalues1, y)
                    critic1_loss.backward()
                    trainer_critic.step(self._minibatch_size)

                    with autograd.record():
                        qvalues2 = self._critic2(states, actions)
                        critic2_loss = l2_loss(qvalues2, y)
                    critic2_loss.backward()
                    trainer_critic2.step(self._minibatch_size)

                    critic_loss = (critic1_loss.mean() + critic2_loss.mean())/2

                    if self._total_steps % self._policy_delay == 0:
                        tmp_critic = self._copy_critic()
                        with autograd.record():
                            actor_loss = -tmp_critic(
                                states, self._actor(states)).mean()
                        actor_loss.backward()
                        trainer_actor.step(self._minibatch_size)

                        # Update target networks:
                        self._actor_target = self._soft_update(
                            self._actor, self._actor_target,
                            self._soft_target_update_rate)
                        self._critic_target = self._soft_update(
                            self._critic, self._critic_target,
                            self._soft_target_update_rate)
                        self._critic2_target = self._soft_update(
                            self._critic2, self._critic2_target,
                            self._soft_target_update_rate)

                        actor_updates = actor_updates + 1
                    else:
                        actor_loss = nd.array([0], ctx=self._ctx)

                    # Update statistics
                    episode_critic_loss +=\
                        np.sum(critic_loss.asnumpy()) / self._minibatch_size
                    episode_actor_loss += 0 if actor_updates == 0 else\
                        np.sum(actor_loss.asnumpy()[0])
                    episode_avg_q_value +=\
                        np.sum(target_qvalues.asnumpy()) / self._minibatch_size

                    training_steps += 1

                episode_reward += reward
                step += 1
                self._total_steps += 1
                state = next_state

                if terminal:
                    # Reset the strategy
                    self._strategy.reset()
                    break

            # Log the episode results
            episode_actor_loss = 0 if actor_updates == 0\
                else (episode_actor_loss / actor_updates)
            episode_critic_loss = 0 if training_steps == 0\
                else (episode_critic_loss / training_steps)
            episode_avg_q_value = 0 if actor_updates == 0\
                else (episode_avg_q_value / training_steps)

            avg_reward = self._training_stats.log_episode(
                self._current_episode, start, training_steps,
                episode_actor_loss, episode_critic_loss, episode_avg_q_value,
                self._strategy.cur_eps, episode_reward)

            self._do_snapshot_if_in_interval(self._current_episode)

            if self._is_target_reached(avg_reward):
                self._logger.info(
                    'Target score is reached in average; Training is stopped')
                break

            self._current_episode += 1

        self._evaluate()
        self.save_parameters(episode=self._current_episode)
        self.export_best_network()
        self._training_stats.save_stats(self._output_directory)
        self._logger.info('--------- Training finished ---------')
        return True

        def _make_config_dict(self):
            config = super(TwinDelayedDdpgAgent, self)._make_config_dict()
            config['policy_noise'] = self._policy_noise
            config['noise_clip'] = self._noise_clip
            config['policy_delay'] = self._policy_delay
            return config


Nicola Gatto's avatar
Nicola Gatto committed
996 997 998 999 1000 1001 1002 1003 1004 1005 1006
class DqnAgent(Agent):
    def __init__(
        self,
        qnet,
        environment,
        replay_memory_params,
        strategy_params,
        state_dim,
        action_dim,
        ctx=None,
        discount_factor=.9,
1007
        loss_function='l2',
Nicola Gatto's avatar
Nicola Gatto committed
1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044
        optimizer='rmsprop',
        optimizer_params={'learning_rate': 0.09},
        training_episodes=50,
        start_training=0,
        train_interval=1,
        use_fix_target=False,
        double_dqn=False,
        target_update_interval=10,
        snapshot_interval=200,
        evaluation_samples=100,
        agent_name='Dqn_agent',
        max_episode_step=99999,
        output_directory='model_parameters',
        verbose=True,
        target_score=None
    ):
        super(DqnAgent, self).__init__(
            environment=environment, replay_memory_params=replay_memory_params,
            strategy_params=strategy_params, state_dim=state_dim,
            action_dim=action_dim, ctx=ctx, discount_factor=discount_factor,
            training_episodes=training_episodes, start_training=start_training,
            train_interval=train_interval,
            snapshot_interval=snapshot_interval, agent_name=agent_name,
            max_episode_step=max_episode_step,
            output_directory=output_directory, verbose=verbose,
            target_score=target_score, evaluation_samples=evaluation_samples)

        self._qnet = qnet
        self._target_update_interval = target_update_interval
        self._target_qnet = copy_net(
            self._qnet, self._state_dim, ctx=self._ctx)
        self._loss_function_str = loss_function
        self._loss_function = get_loss_function(loss_function)
        self._optimizer = optimizer
        self._optimizer_params = optimizer_params
        self._double_dqn = double_dqn
        self._use_fix_target = use_fix_target
1045

1046 1047 1048 1049 1050 1051 1052 1053
        # Build memory buffer for discrete actions
        replay_memory_params['state_dim'] = state_dim
        replay_memory_params['action_dim'] = (1,)
        self._replay_memory_params = replay_memory_params
        rm_builder = ReplayMemoryBuilder()
        self._memory = rm_builder.build_by_params(**replay_memory_params)
        self._minibatch_size = self._memory.sample_size

Nicola Gatto's avatar
Nicola Gatto committed
1054 1055 1056
        # Initialize best network
        self._best_net = copy_net(self._qnet, self._state_dim, self._ctx)
        self._best_avg_score = -np.infty
1057

Nicola Gatto's avatar
Nicola Gatto committed
1058
        self._training_stats = None
1059

Nicola Gatto's avatar
Nicola Gatto committed
1060 1061 1062 1063 1064
    @classmethod
    def resume_from_session(cls, session_dir, net, environment):
        import pickle
        if not os.path.exists(session_dir):
            raise ValueError('Session directory does not exist')
1065

Nicola Gatto's avatar
Nicola Gatto committed
1066 1067 1068 1069 1070 1071
        files = dict()
        files['agent'] = os.path.join(session_dir, 'agent.p')
        files['best_net_params'] = os.path.join(session_dir, 'best_net.params')
        files['q_net_params'] = os.path.join(session_dir, 'qnet.params')
        files['target_net_params'] = os.path.join(
            session_dir, 'target_net.params')
1072

Nicola Gatto's avatar
Nicola Gatto committed
1073 1074 1075 1076 1077
        for file in files.values():
            if not os.path.exists(file):
                raise ValueError(
                    'Session directory is not complete: {} is missing'
                    .format(file))
1078

Nicola Gatto's avatar
Nicola Gatto committed
1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
        with open(files['agent'], 'rb') as f:
            agent = pickle.load(f)

        agent._environment = environment
        agent._qnet = net
        agent._qnet.load_parameters(files['q_net_params'], agent._ctx)
        agent._qnet.hybridize()
        agent._qnet(nd.random_normal(
            shape=((1,) + agent._state_dim), ctx=agent._ctx))
        agent._best_net = copy_net(agent._qnet, agent._state_dim, agent._ctx)
        agent._best_net.load_parameters(files['best_net_params'], agent._ctx)
        agent._target_qnet = copy_net(
            agent._qnet, agent._state_dim, agent._ctx)
        agent._target_qnet.load_parameters(
            files['target_net_params'], agent._ctx)

        agent._logger = ArchLogger.get_logger()
        agent._training_stats.logger = agent._logger
        agent._logger.info('Agent was retrieved; Training can be continued')

        return agent

    def _make_pickle_ready(self, session_dir):
        super(DqnAgent, self)._make_pickle_ready(session_dir)
1103 1104
        self._export_net(self._qnet, 'current_qnet')
        self._export_net(self._qnet, 'qnet', filedir=session_dir)
Nicola Gatto's avatar
Nicola Gatto committed
1105
        self._qnet = None
1106
        self._export_net(self._target_qnet, 'target_net', filedir=session_dir)
Nicola Gatto's avatar
Nicola Gatto committed
1107
        self._target_qnet = None
1108 1109

    def get_q_values(self, state, with_best=False):
Nicola Gatto's avatar
Nicola Gatto committed
1110 1111
        return self.get_batch_q_values(
            nd.array([state], ctx=self._ctx), with_best=with_best)[0]
1112 1113

    def get_batch_q_values(self, state_batch, with_best=False):
Nicola Gatto's avatar
Nicola Gatto committed
1114 1115
        return self._best_net(state_batch)\
            if with_best else self._qnet(state_batch)
1116 1117 1118 1119

    def get_next_action(self, state, with_best=False):
        q_values = self.get_q_values(state, with_best=with_best)
        action = q_values.asnumpy().argmax()
Nicola Gatto's avatar
Nicola Gatto committed
1120
        return action
1121

Nicola Gatto's avatar
Nicola Gatto committed
1122 1123 1124 1125 1126
    def __determine_target_q_values(
        self, states, actions, rewards, next_states, terminals
    ):
        if self._use_fix_target:
            q_max_val = self._target_qnet(next_states)
1127
        else:
Nicola Gatto's avatar
Nicola Gatto committed
1128
            q_max_val = self._qnet(next_states)
1129

Nicola Gatto's avatar
Nicola Gatto committed
1130 1131 1132 1133 1134
        if self._double_dqn:
            q_values_next_states = self._qnet(next_states)
            target_rewards = rewards + nd.choose_element_0index(
                q_max_val, nd.argmax_channel(q_values_next_states))\
                * (1.0 - terminals) * self._discount_factor
1135
        else:
Nicola Gatto's avatar
Nicola Gatto committed
1136 1137 1138
            target_rewards = rewards + nd.choose_element_0index(
                q_max_val, nd.argmax_channel(q_max_val))\
                * (1.0 - terminals) * self._discount_factor
1139

Nicola Gatto's avatar
Nicola Gatto committed
1140
        target_qval = self._qnet(states)
1141 1142 1143 1144 1145 1146
        for t in range(target_rewards.shape[0]):
            target_qval[t][actions[t]] = target_rewards[t]

        return target_qval

    def __train_q_net_step(self, trainer):
Nicola Gatto's avatar
Nicola Gatto committed
1147 1148 1149 1150
        states, actions, rewards, next_states, terminals =\
            self._sample_from_memory()
        target_qval = self.__determine_target_q_values(
            states, actions, rewards, next_states, terminals)
1151
        with autograd.record():
Nicola Gatto's avatar
Nicola Gatto committed
1152 1153
            q_values = self._qnet(states)
            loss = self._loss_function(q_values, target_qval)
1154
        loss.backward()
Nicola Gatto's avatar
Nicola Gatto committed
1155
        trainer.step(self._minibatch_size)
1156 1157 1158
        return loss

    def __do_target_update_if_in_interval(self, total_steps):
Nicola Gatto's avatar
Nicola Gatto committed
1159 1160 1161
        do_target_update = (
            self._use_fix_target and
            (total_steps % self._target_update_interval == 0))
1162
        if do_target_update:
Nicola Gatto's avatar
Nicola Gatto committed
1163 1164 1165 1166
            self._logger.info(
                'Target network is updated after {} steps'.format(total_steps))
            self._target_qnet = copy_net(
                self._qnet, self._state_dim, self._ctx)
1167 1168

    def train(self, episodes=None):
Nicola Gatto's avatar
Nicola Gatto committed
1169 1170 1171 1172 1173 1174 1175 1176 1177 1178
        self.save_config_file()
        self._logger.info("--- Start training ---")
        trainer = gluon.Trainer(
            self._qnet.collect_params(),
            self._optimizer,
            self._adjust_optimizer_params(self._optimizer_params))
        episodes = episodes if episodes is not None\
            else self._training_episodes

        resume = (self._current_episode > 0)
1179
        if resume:
Nicola Gatto's avatar
Nicola Gatto committed
1180 1181 1182
            self._logger.info("Training session resumed")
            self._logger.info("Starting from episode {}".format(
                self._current_episode))
1183
        else:
Nicola Gatto's avatar
Nicola Gatto committed
1184
            self._training_stats = DqnTrainingStats(episodes)
1185

Nicola Gatto's avatar
Nicola Gatto committed
1186 1187 1188 1189
        # Implementation Deep Q Learning described by
        # Mnih et. al. in Playing Atari with Deep Reinforcement Learning
        while self._current_episode < episodes:
            if self._check_interrupt_routine():
1190 1191 1192 1193 1194
                return False

            step = 0
            episode_reward = 0
            start = time.time()
Nicola Gatto's avatar
Nicola Gatto committed
1195
            state = self._environment.reset()
1196 1197
            episode_loss = 0
            training_steps = 0
Nicola Gatto's avatar
Nicola Gatto committed
1198 1199 1200 1201
            while step < self._max_episode_step:
                # 1. Choose an action based on current game state and policy
                q_values = self._qnet(nd.array([state], ctx=self._ctx))
                action = self._strategy.select_action(q_values[0])
1202
                self._strategy.decay(self._current_episode)
1203

Nicola Gatto's avatar
Nicola Gatto committed
1204 1205 1206
                # 2. Play the game for a single step
                next_state, reward, terminal, _ =\
                    self._environment.step(action)
1207

Nicola Gatto's avatar
Nicola Gatto committed
1208 1209
                # 3. Store transition in replay memory
                self._memory.append(
1210
                    state, [action], reward, next_state, terminal)
1211

Nicola Gatto's avatar
Nicola Gatto committed
1212 1213
                # 4. Train the network if in interval
                if self._do_training():
1214 1215
                    loss = self.__train_q_net_step(trainer)
                    training_steps += 1
Nicola Gatto's avatar
Nicola Gatto committed
1216 1217
                    episode_loss +=\
                        np.sum(loss.asnumpy()) / self._minibatch_size
1218 1219

                # Update target network if in interval
Nicola Gatto's avatar
Nicola Gatto committed
1220
                self.__do_target_update_if_in_interval(self._total_steps)
1221 1222

                step += 1
Nicola Gatto's avatar
Nicola Gatto committed
1223
                self._total_steps += 1
1224 1225 1226 1227
                episode_reward += reward
                state = next_state

                if terminal:
Nicola Gatto's avatar
Nicola Gatto committed
1228
                    self._strategy.reset()
1229 1230
                    break

Nicola Gatto's avatar
Nicola Gatto committed
1231
            self._do_snapshot_if_in_interval(self._current_episode)
1232

Nicola Gatto's avatar
Nicola Gatto committed
1233 1234 1235 1236 1237
            episode_loss = (episode_loss / training_steps)\
                if training_steps > 0 else 0
            avg_reward = self._training_stats.log_episode(
                self._current_episode, start, training_steps,
                episode_loss, self._strategy.cur_eps, episode_reward)
1238

Nicola Gatto's avatar
Nicola Gatto committed
1239 1240 1241
            if self._is_target_reached(avg_reward):
                self._logger.info(
                    'Target score is reached in average; Training is stopped')
1242 1243
                break

Nicola Gatto's avatar
Nicola Gatto committed
1244
            self._current_episode += 1
1245

Nicola Gatto's avatar
Nicola Gatto committed
1246 1247
        self._evaluate()
        self.save_parameters(episode=self._current_episode)
1248
        self.export_best_network()
Nicola Gatto's avatar
Nicola Gatto committed
1249 1250 1251
        self._training_stats.save_stats(self._output_directory)
        self._logger.info('--------- Training finished ---------')
        return True
1252

Nicola Gatto's avatar
Nicola Gatto committed
1253 1254 1255 1256 1257 1258 1259 1260
    def _make_config_dict(self):
        config = super(DqnAgent, self)._make_config_dict()
        config['optimizer'] = self._optimizer
        config['optimizer_params'] = self._optimizer_params
        config['loss_function'] = self._loss_function_str
        config['use_fix_target'] = self._use_fix_target
        config['double_dqn'] = self._double_dqn
        config['target_update_interval'] = self._target_update_interval
1261 1262
        return config

Nicola Gatto's avatar
Nicola Gatto committed
1263
    def save_parameters(self, episode):
1264 1265
        self._export_net(
            self._qnet, self._agent_name + '_qnet', episode=episode)
Nicola Gatto's avatar
Nicola Gatto committed
1266 1267 1268

    def _save_current_as_best_net(self):
        self._best_net = copy_net(
1269
            self._qnet, self._state_dim, ctx=self._ctx)