Merge branch 'first_version' into 'master'

version 1.0.0

See merge request !1

README.md

-We need some helpful information here
\ No newline at end of file

README.rst

+WiseMove is safe reinforcement learning framework that combines hierarchical reinforcement learning and model-checking using temporal logic constraints.
+
+Requirements
+============
+
+* Python 3.6
+* `requirements.txt <../requirements.txt>`_ for package list.
+
+
+Installation
+============
+
+* Run the install script: :code:`scripts/install.sh`
+
+Documentation
+=============
+
+* For documentation, we use *Sphinx*
+    * If `install.sh` was run, use :code:`scripts/generate_doc.sh launch` to view documentation or open `index.html <./documentation/index.html>`_ .
+    * If not, use :code:`scripts/generate_doc.sh build` to generate documentation first.
+
+Replicate Results
+=================
+These are the minimum steps required to replicate the results for simple_intersection environment. For a detailed user guide, it is recommended to view the documentation.
+
+* Run `scripts/install.sh`
+* Low-level policies:
+    * To train all low-level policies from scratch: `python low_level_policy_main.py --train`
+    * To train a single low-level, for example wait: `python low_level_policy_main.py --option=wait --train`
+    * To test these trained low-level policies: `python low_level_policy_main.py --test --saved_policy_in_root`
+    * To test one of these trained low-level policies, for example wait: `python low_level_policy_main.py --option=wait --test --saved_policy_in_root`
+* High-level policy:
+    * To train high-level policy from scratch using the given low-level policies: `python high_level_policy_main.py --train`
+    * To evaluate this trained high-level policy: `python high_level_policy_main.py --evaluate --saved_policy_in_root`
+* To run MCTS using the high-level policy: `python mcts.py`
+
+Coding Standards
+================
+
+We follow PEP8 style guidelines for coding and PEP257 for documentation.
+It is not necessary to keep these in mind while coding, but before
+submitting a pull request, do these two steps for each python file you 
+have modified.
+
+1. :code:`yapf -i YOUR_MODIFIED_FILE.py`
+2. :code:`docformatter --in-place YOUR_MODIFIED_FILE.py`
+
+:code:`yapf` formats the code and :code:`docformatter` formats the docstrings.
\ No newline at end of file

backends/__init__.py

+from .manual_policy import ManualPolicy
+from .mcts_learner import MCTSLearner
+from .rl_controller import RLController
+from .kerasrl_learner import DDPGLearner, DQNLearner
+from .online_mcts_controller import OnlineMCTSController
\ No newline at end of file

backends/baselines_learner.py

+from .learner_base import LearnerBase
+
+# TODO: make sure that the package for PPO2 is installed.
+from stable_baselines import PPO2
+from stable_baselines.common.vec_env import DummyVecEnv
+from stable_baselines.common.policies import MlpPolicy
+
+import numpy as np
+
+
+class PPO2Agent(LearnerBase):
+    def __init__(self,
+                 input_shape,
+                 nb_actions,
+                 env,
+                 policy=None,
+                 tensorboard=False,
+                 log_path="./logs",
+                 **kwargs):
+        """The constructor which sets the properties of the class.
+
+        Args:
+            input_shape: Shape of observation space, e.g (10,);
+            nb_actions: number of values in action space;
+            env: env on which the agent learns
+            policy: stable_baselines Policy object. default is MlpPolicy
+            tensorboard: whether to integrate tensorboard or not
+            log_path="./logs",
+            **kwargs: other optional key-value arguments with defaults defined in property_defaults
+        """
+        super(PPO2Agent, self).__init__(input_shape, nb_actions, **kwargs)
+
+        if policy is None:
+            policy = self.get_default_policy()
+
+        self.log_path = log_path
+
+        self.env = DummyVecEnv([lambda: env]) #PPO2 requried a vectorized environment for parallel training
+        self.agent_model = self.create_agent(policy, tensorboard)
+
+    def get_default_policy(self):
+        """Creates the default policy.
+
+        Returns:     stable_baselines Policy object. default is MlpPolicy
+        """
+        return MlpPolicy
+
+    def create_agent(self, policy, tensorboard):
+        """Creates a PPO agent
+
+        Returns:
+            stable_baselines PPO2 object
+        """
+        if tensorboard:
+            return PPO2(policy, self.env, verbose=1, tensorboard_log=self.log_path)
+        else:
+            return PPO2(policy, self.env, verbose=1)
+
+    def fit(self,
+            env=None,
+            nb_steps=1000000,
+            visualize=False,
+            nb_max_episode_steps=200):
+
+        # PPO2 callback is only called each episode (not step) so cannot render whole episode
+        # To render each step, add self.env.render() in at Runner class method run() in stable_baselines ppo2.py
+        callback = self.__render_env_while_learning if visualize else None
+        self.agent_model.learn(total_timesteps=nb_steps, callback=callback)
+
+    @staticmethod
+    def __render_env_while_learning(_locals, _globals):
+        _locals['self'].env.render()
+
+    def save_weights(self, file_name="test_weights.h5f", overwrite=True):
+        self.agent_model.save(file_name)
+
+    def test_model(self,
+                   env=None,
+                   nb_episodes=50,
+                   visualize=True,
+                   nb_max_episode_steps=200):
+
+        episode_rewards = [0.0]
+        obs = self.env.reset()
+        current_episode = 1
+        current_step = 0
+        while current_episode <= nb_episodes:
+            # _states are only useful when using LSTM policies
+            action, _states = self.agent_model.predict(obs)
+
+            # here, action, rewards  and dones are arrays
+            # because we are using vectorized env
+            obs, rewards, dones, info = self.env.step(action)
+            current_step += 1
+
+            if visualize:
+                self.env.render()
+
+            # Stats
+            episode_rewards[-1] += rewards[0]
+            if dones[0] or current_step > nb_max_episode_steps:
+                obs = self.env.reset()
+                print ("Episode ", current_episode, "reward: ", episode_rewards[-1])
+                episode_rewards.append(0.0)
+                current_episode += 1
+                current_step = 0
+
+        # Compute mean reward for the last 100 episodes
+        mean_100ep_reward = round(np.mean(episode_rewards[-100:]), 1)
+        print("Mean reward over last 100 episodes:", mean_100ep_reward)
+
+    def load_weights(self, file_name="test_weights.h5f"):
+        self.agent_model = PPO2.load(file_name)
+
+    def forward(self, observation):
+        return self.agent_model.predict(observation)
+
+    def set_environment(self, env):
+        self.env = DummyVecEnv([lambda: env])
+        self.agent_model.set_env(self.env)

backends/controller_base.py

+from .policy_base import PolicyBase
+
+
+class ControllerBase(PolicyBase):
+    """Abstract class for controllers."""
+
+    def __init__(self, env, low_level_policies, start_node_alias):
+        self.env = env
+        self.low_level_policies = low_level_policies
+
+        # TODO: Move an intermediate class so that base class can be clean
+        self.current_node = self.low_level_policies[start_node_alias]
+        self.node_terminal_state_reached = False
+        self.controller_args_defaults = {}
+
+    def set_controller_args(self, **kwargs):
+        for (prop, default) in self.controller_args_defaults.items():
+            setattr(self, prop, kwargs.get(prop, default))
+
+    def can_transition(self):
+        """Returns boolean signifying whether we can transition. To be
+        implemented in subclass.
+        """
+
+        raise NotImplemented(self.__class__.__name__ + \
+            "can_transition is not implemented.")
+
+    def do_transition(self, observation):
+        """Do a transition, assuming we can transition. To be
+        implemented in subclass.
+
+        Args:
+            observation: final observation from episodic step
+        """
+
+        raise NotImplemented(self.__class__.__name__ + \
+            "do_transition is not implemented.")
+
+    def set_current_node(self, node_alias):
+        """Sets the current node which is being executed
+
+        Args:
+            node: node alias of the node to be set
+        """
+        raise NotImplemented(self.__class__.__name__ + \
+            "set_current_node is not implemented.")
+
+    # TODO: Looks generic. Move to an intermediate class/highlevel manager so that base class can be clean
+    ''' Executes the current node until node termination condition is reached
+    
+    Returns state at end of node execution, total reward, epsiode_termination_flag, info
+    '''
+    def step_current_node(self, visualize_low_level_steps=False):
+        total_reward = 0
+        self.node_terminal_state_reached = False
+        while not self.node_terminal_state_reached:
+            observation, reward, terminal, info = self.low_level_step_current_node()
+            if visualize_low_level_steps:
+                self.env.render()
+            total_reward += reward
+
+        total_reward += self.current_node.high_level_extra_reward
+
+        # TODO for info
+        return observation, total_reward, self.env.termination_condition, info
+
+    # TODO: Looks generic. Move to an intermediate class/highlevel manager so that base class can be clean
+    ''' Executes one step of current node. Sets node_terminal_state_reached flag if node termination condition
+    has been reached. 
+
+    Returns state after one step, step reward, episode_termination_flag, info
+    '''
+    def low_level_step_current_node(self):
+
+        u_ego = self.current_node.low_level_policy(self.current_node.get_reduced_features_tuple())
+        feature, R, terminal, info = self.current_node.step(u_ego)
+        self.node_terminal_state_reached = terminal
+        return self.env.get_features_tuple(), R, self.env.termination_condition, info

backends/kerasrl_learner.py

+from .learner_base import LearnerBase
+
+from keras.models import Sequential, Model
+from keras.layers import Dense, Activation, Flatten, Input, Concatenate
+from keras.optimizers import Adam
+from keras.callbacks import TensorBoard
+
+from rl.agents import DDPGAgent, DQNAgent
+from rl.memory import SequentialMemory
+from rl.random import OrnsteinUhlenbeckProcess
+from rl.policy import BoltzmannQPolicy, MaxBoltzmannQPolicy
+
+from rl.callbacks import ModelIntervalCheckpoint
+
+import numpy as np
+
+
+class DDPGLearner(LearnerBase):
+    def __init__(self,
+                 input_shape=(48, ),
+                 nb_actions=2,
+                 actor=None,
+                 critic=None,
+                 critic_action_input=None,
+                 memory=None,
+                 random_process=None,
+                 **kwargs):
+        """The constructor which sets the properties of the class.
+
+        Args:
+            input_shape: Shape of observation space, e.g (10,);
+            nb_actions: number of values in action space;
+            actor: Keras Model of actor which takes observation as input and outputs actions. Uses default if not given
+            critic: Keras Model of critic which takes concatenation of observation and action and outputs a single
+                value. Uses default if not given
+            critic_action_input: Keras Input which was used in creating action input of the critic model.
+                Uses default critic and action_input if not specified
+            memory: KerasRL Memory. Uses default SequentialMemory if not given
+            random_process: KerasRL random process. Uses default OrnsteinUhlenbeckProcess if not given
+            **kwargs: other optional key-value arguments with defaults defined in property_defaults
+        """
+        super(DDPGLearner, self).__init__(input_shape, nb_actions, **kwargs)
+        property_defaults = {
+            "mem_size": 100000,  # size of memory
+            "mem_window_length": 1,  # window length of memory
+            "oup_theta": 0.15,  # OrnsteinUhlenbeckProcess theta
+            "oup_mu": 0,  # OrnsteinUhlenbeckProcess mu
+            "oup_sigma": 1,  # OrnsteinUhlenbeckProcess sigma
+            "oup_sigma_min": 0.5,  # OrnsteinUhlenbeckProcess sigma min
+            "oup_annealing_steps": 500000, # OrnsteinUhlenbeckProcess n-step annealing
+            "nb_steps_warmup_critic": 100,  # steps for critic to warmup
+            "nb_steps_warmup_actor": 100,  # steps for actor to warmup
+            "target_model_update": 1e-3  # target model update frequency
+        }
+
+        for (prop, default) in property_defaults.items():
+            setattr(self, prop, kwargs.get(prop, default))
+
+        if actor is None:
+            actor = self.get_default_actor_model()
+        if critic is None or critic_action_input is None:
+            critic, critic_action_input = self.get_default_critic_model()
+        if memory is None:
+            memory = self.get_default_memory()
+        if random_process is None:
+            random_process = self.get_default_randomprocess()
+
+        #TODO: Add output scaling
+        self.agent_model = self.create_agent(
+            actor, critic, critic_action_input, memory, random_process)
+
+    def get_default_actor_model(self):
+        """Creates the default actor model.
+
+        Returns:     Keras Model object of actor
+        """
+        actor = Sequential()
+        actor.add(Flatten(input_shape=(1, ) + self.input_shape))
+        actor.add(Dense(64, use_bias=False))
+        actor.add(Activation('relu'))
+        actor.add(Dense(64, use_bias=False))
+        actor.add(Activation('relu'))
+        actor.add(Dense(self.nb_actions, use_bias=True))
+        actor.add(Activation('tanh'))
+
+        # print(actor.summary())
+
+        return actor
+
+    def get_default_critic_model(self):
+        """Creates the default critic model.
+
+        Returns:     Keras Model object of critic
+        """
+        action_input = Input(shape=(self.nb_actions, ), name='action_input')
+        observation_input = Input(
+            shape=(1, ) + self.input_shape, name='observation_input')
+        flattened_observation = Flatten()(observation_input)
+        x = Concatenate()([action_input, flattened_observation])
+        x = Dense(64, use_bias=False)(x)
+        x = Activation('relu')(x)
+        x = Dense(64, use_bias=False)(x)
+        x = Activation('relu')(x)
+        x = Dense(64, use_bias=False)(x)
+        x = Activation('relu')(x)
+        x = Dense(1, use_bias=True)(x)
+        #x = Activation('linear')(x)
+        critic = Model(inputs=[action_input, observation_input], outputs=x)
+        # print(critic.summary())
+
+        return critic, action_input
+
+    def get_default_randomprocess(self):
+        """Creates the default random process model.
+
+        Returns:     KerasRL OrnsteinUhlenbeckProcess object
+        """
+        random_process = OrnsteinUhlenbeckProcess(
+            size=self.nb_actions,
+            theta=self.oup_theta,
+            mu=self.oup_mu,
+            sigma=self.oup_sigma,
+            sigma_min=self.oup_sigma_min,
+            n_steps_annealing=self.oup_annealing_steps)
+        return random_process
+
+    def get_default_memory(self):
+        """Creates the default memory model.
+
+        Returns:     KerasRL SequentialMemory object
+        """
+        memory = SequentialMemory(
+            limit=self.mem_size, window_length=self.mem_window_length)
+        return memory
+
+    def create_agent(self, actor, critic, critic_action_input, memory,
+                     random_process):
+        """Creates a KerasRL DDPGAgent with given components.
+
+        Args:
+            actor: Keras Model of actor which takes observation as input and outputs actions.
+            critic: Keras Model of critic that takes concatenation of observation and action and outputs a single value.
+            critic_action_input: Keras Input which was used in creating action input of the critic model.
+            memory: KerasRL Memory.
+            random_process: KerasRL random process.
+
+        Returns:
+            KerasRL DDPGAgent object
+        """
+        agent = DDPGAgent(
+            nb_actions=self.nb_actions,
+            actor=actor,
+            critic=critic,
+            critic_action_input=critic_action_input,
+            memory=memory,
+            nb_steps_warmup_critic=self.nb_steps_warmup_critic,
+            nb_steps_warmup_actor=self.nb_steps_warmup_actor,
+            random_process=random_process,
+            gamma=self.gamma,
+            target_model_update=1e-3)
+
+        # TODO: give params like lr_actor and lr_critic to set different lr of Actor and Critic.
+        agent.compile([Adam(lr=self.lr*1e-2, clipnorm=1.), Adam(lr=self.lr, clipnorm=1.)], metrics=['mae'])
+
+        return agent
+
+    def train(self,
+            env,
+            nb_steps=1000000,
+            visualize=False,
+            verbose=1,
+            log_interval=10000,
+            nb_max_episode_steps=200,
+            model_checkpoints=False,
+            checkpoint_interval=100000,
+            tensorboard=False):
+
+        callbacks = []
+        if model_checkpoints:
+            callbacks += [ModelIntervalCheckpoint('./checkpoints/checkpoint_weights.h5f', interval=checkpoint_interval)]
+        if tensorboard:
+            callbacks += [TensorBoard(log_dir='./logs')]
+
+        self.agent_model.fit(
+            env,
+            nb_steps=nb_steps,
+            visualize=visualize,
+            verbose=verbose,
+            log_interval=log_interval,
+            nb_max_episode_steps=nb_max_episode_steps,
+            callbacks=callbacks)
+
+    def save_model(self, file_name="test_weights.h5f", overwrite=True):
+        self.agent_model.save_weights(file_name, overwrite=True)
+
+    def test_model(self,
+                   env,
+                   nb_episodes=50,
+                   visualize=True,
+                   nb_max_episode_steps=200):
+        self.agent_model.test(
+            env,
+            nb_episodes=nb_episodes,
+            visualize=visualize,
+            nb_max_episode_steps=nb_max_episode_steps)
+
+    def load_model(self, file_name="test_weights.h5f"):
+        self.agent_model.load_weights(file_name)
+
+    def predict(self, observation):
+        return self.agent_model.forward(observation)
+
+
+class DQNLearner(LearnerBase):
+    def __init__(self,
+                 input_shape=(48, ),
+                 nb_actions=5,
+                 low_level_policies=None,
+                 model=None,
+                 policy=None,
+                 memory=None,
+                 **kwargs):
+        """The constructor which sets the properties of the class.
+
+        Args:
+            input_shape: Shape of observation space, e.g (10,);
+            nb_actions: number of values in action space;
+            model: Keras Model of actor which takes observation as input and outputs actions. Uses default if not given
+            policy: KerasRL Policy. Uses default SequentialMemory if not given
+            memory: KerasRL Memory. Uses default BoltzmannQPolicy if not given
+            **kwargs: other optional key-value arguments with defaults defined in property_defaults
+        """
+        super(DQNLearner, self).__init__(input_shape, nb_actions, **kwargs)
+        property_defaults = {
+            "mem_size": 100000,  # size of memory
+            "mem_window_length": 1,  # window length of memory
+            "target_model_update": 1e-3,  # target model update frequency
+            "nb_steps_warmup": 100,  # steps for model to warmup
+        }
+
+        for (prop, default) in property_defaults.items():
+            setattr(self, prop, kwargs.get(prop, default))
+
+        if model is None:
+            model = self.get_default_model()
+        if policy is None:
+            policy = self.get_default_policy()
+        if memory is None:
+            memory = self.get_default_memory()
+
+        self.low_level_policies = low_level_policies
+
+        self.agent_model = self.create_agent(model, policy, memory)
+
+    def get_default_model(self):
+        """Creates the default model.
+
+        Returns:     Keras Model object of actor
+        """
+        model = Sequential()
+        model.add(Flatten(input_shape=(1, ) + self.input_shape))
+        model.add(Dense(32))
+        model.add(Activation('relu'))
+        model.add(Dense(32))
+        model.add(Activation('relu'))
+        model.add(Dense(self.nb_actions))
+        model.add(Activation('linear'))
+        # print(model.summary())
+
+        return model
+
+    def get_default_policy(self):
+        return MaxBoltzmannQPolicy(eps=0.3)
+
+    def get_default_memory(self):
+        """Creates the default memory model.
+
+        Returns:     KerasRL SequentialMemory object
+        """
+        memory = SequentialMemory(
+            limit=self.mem_size, window_length=self.mem_window_length)
+        return memory
+
+    def create_agent(self, model, policy, memory):
+        """Creates a KerasRL DDPGAgent with given components.
+
+        Args:
+            model: Keras Model of model which takes observation as input and outputs discrete actions.
+            memory: KerasRL Memory.
+
+        Returns:
+            KerasRL DQN object
+        """
+        agent = DQNAgentOverOptions(model=model, low_level_policies=self.low_level_policies,
+                                         nb_actions=self.nb_actions, memory=memory,
+                                         nb_steps_warmup=self.nb_steps_warmup, target_model_update=self.target_model_update,
+                                         policy=policy, enable_dueling_network=True)
+
+
+        agent.compile(Adam(lr=self.lr), metrics=['mae'])
+