coach v0.8.0

exploration_policies/__init__.py

@@ -0,0 +1,28 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.additive_noise import *
+from exploration_policies.approximated_thompson_sampling_using_dropout import *
+from exploration_policies.bayesian import *
+from exploration_policies.boltzmann import *
+from exploration_policies.bootstrapped import *
+from exploration_policies.categorical import *
+from exploration_policies.continuous_entropy import *
+from exploration_policies.e_greedy import *
+from exploration_policies.exploration_policy import *
+from exploration_policies.greedy import *
+from exploration_policies.ou_process import *
+from exploration_policies.thompson_sampling import *

exploration_policies/additive_noise.py

@@ -0,0 +1,46 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import numpy as np
+from exploration_policies.exploration_policy import *
+
+
+class AdditiveNoise(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.variance = tuning_parameters.exploration.initial_noise_variance_percentage
+        self.final_variance = tuning_parameters.exploration.final_noise_variance_percentage
+        self.decay_steps = tuning_parameters.exploration.noise_variance_decay_steps
+        self.variance_decay_delta = (self.variance - self.final_variance) / float(self.decay_steps)
+
+    def decay_exploration(self):
+        if self.variance > self.final_variance:
+            self.variance -= self.variance_decay_delta
+        elif self.variance < self.final_variance:
+            self.variance = self.final_variance
+
+    def get_action(self, action_values):
+        if self.phase == RunPhase.TRAIN:
+            self.decay_exploration()
+        action = np.random.normal(action_values, 2 * self.variance * self.action_abs_range)
+        return action #np.clip(action, -self.action_abs_range, self.action_abs_range).squeeze()
+
+    def get_control_param(self):
+        return self.variance

exploration_policies/approximated_thompson_sampling_using_dropout.py

@@ -0,0 +1,40 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+
+
+class ApproximatedThompsonSamplingUsingDropout(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.dropout_discard_probability = tuning_parameters.exploration.dropout_discard_probability
+        self.network = tuning_parameters.network
+        self.assign_op = self.network.dropout_discard_probability.assign(self.dropout_discard_probability)
+        self.network.sess.run(self.assign_op)
+        pass
+
+    def decay_dropout(self):
+        pass
+
+    def get_action(self, action_values):
+        return np.argmax(action_values)
+
+    def get_control_param(self):
+        return self.dropout_discard_probability

exploration_policies/bayesian.py

@@ -0,0 +1,56 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+import tensorflow as tf
+
+
+class Bayesian(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.keep_probability = tuning_parameters.exploration.initial_keep_probability
+        self.final_keep_probability = tuning_parameters.exploration.final_keep_probability
+        self.keep_probability_decay_delta = (
+                                            tuning_parameters.exploration.initial_keep_probability - tuning_parameters.exploration.final_keep_probability) \
+                                            / float(tuning_parameters.exploration.keep_probability_decay_steps)
+        self.action_space_size = tuning_parameters.env.action_space_size
+        self.network = tuning_parameters.network
+        self.epsilon = 0
+
+    def decay_keep_probability(self):
+        if (self.keep_probability > self.final_keep_probability and self.keep_probability_decay_delta > 0) \
+                or (self.keep_probability < self.final_keep_probability and self.keep_probability_decay_delta < 0):
+            self.keep_probability -= self.keep_probability_decay_delta
+
+    def get_action(self, action_values):
+        if self.phase == RunPhase.TRAIN:
+            self.decay_keep_probability()
+        # dropout = self.network.get_layer('variable_dropout_1')
+        # with tf.Session() as sess:
+        #     print(dropout.rate.eval())
+        # set_value(dropout.rate, 1-self.keep_probability)
+
+        print(self.keep_probability)
+        self.network.curr_keep_prob = self.keep_probability
+
+        return np.argmax(action_values)
+
+    def get_control_param(self):
+        return self.keep_probability

exploration_policies/boltzmann.py

@@ -0,0 +1,48 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+
+
+class Boltzmann(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.temperature = tuning_parameters.exploration.initial_temperature
+        self.final_temperature = tuning_parameters.exploration.final_temperature
+        self.temperature_decay_delta = (
+                                       tuning_parameters.exploration.initial_temperature - tuning_parameters.exploration.final_temperature) \
+                                       / float(tuning_parameters.exploration.temperature_decay_steps)
+
+    def decay_temperature(self):
+        if self.temperature > self.final_temperature:
+            self.temperature -= self.temperature_decay_delta
+
+    def get_action(self, action_values):
+        if self.phase == RunPhase.TRAIN:
+            self.decay_temperature()
+        # softmax calculation
+        exp_probabilities = np.exp(action_values / self.temperature)
+        probabilities = exp_probabilities / np.sum(exp_probabilities)
+        probabilities[-1] = 1 - np.sum(probabilities[:-1])  # make sure probs sum to 1
+        # choose actions according to the probabilities
+        return np.random.choice(range(self.action_space_size), p=probabilities)
+
+    def get_control_param(self):
+        return self.temperature

exploration_policies/bootstrapped.py

@@ -0,0 +1,37 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.e_greedy import *
+
+
+class Bootstrapped(EGreedy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running parameters
+        :type tuning_parameters: Preset
+        """
+        EGreedy.__init__(self, tuning_parameters)
+        self.num_heads = tuning_parameters.exploration.architecture_num_q_heads
+        self.selected_head = 0
+
+    def select_head(self):
+        self.selected_head = np.random.randint(self.num_heads)
+
+    def get_action(self, action_values):
+        return EGreedy.get_action(self, action_values[self.selected_head])
+
+    def get_control_param(self):
+        return self.selected_head

exploration_policies/categorical.py

@@ -0,0 +1,33 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+
+
+class Categorical(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+
+    def get_action(self, action_values):
+        # choose actions according to the probabilities
+        return np.random.choice(range(self.action_space_size), p=action_values)
+
+    def get_control_param(self):
+        return 0

exploration_policies/continuous_entropy.py

@@ -0,0 +1,22 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import numpy as np
+from exploration_policies.exploration_policy import *
+
+
+class ContinuousEntropy(ExplorationPolicy):
+    pass

exploration_policies/e_greedy.py

@@ -0,0 +1,70 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+
+
+class EGreedy(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.epsilon = tuning_parameters.exploration.initial_epsilon
+        self.final_epsilon = tuning_parameters.exploration.final_epsilon
+        self.epsilon_decay_delta = (
+                                   tuning_parameters.exploration.initial_epsilon - tuning_parameters.exploration.final_epsilon) \
+                                   / float(tuning_parameters.exploration.epsilon_decay_steps)
+        self.evaluation_epsilon = tuning_parameters.exploration.evaluation_epsilon
+
+        # for continuous e-greedy (see http://www.cs.ubc.ca/~van/papers/2017-TOG-deepLoco/2017-TOG-deepLoco.pdf)
+        self.variance = tuning_parameters.exploration.initial_noise_variance_percentage
+        self.final_variance = tuning_parameters.exploration.final_noise_variance_percentage
+        self.decay_steps = tuning_parameters.exploration.noise_variance_decay_steps
+        self.variance_decay_delta = (self.variance - self.final_variance) / float(self.decay_steps)
+
+    def decay_exploration(self):
+        # decay epsilon
+        if self.epsilon > self.final_epsilon:
+            self.epsilon -= self.epsilon_decay_delta
+        elif self.epsilon < self.final_epsilon:
+            self.epsilon = self.final_epsilon
+
+        # decay noise variance
+        if not self.discrete_controls:
+            if self.variance > self.final_variance:
+                self.variance -= self.variance_decay_delta
+            elif self.variance < self.final_variance:
+                self.variance = self.final_variance
+
+    def get_action(self, action_values):
+        if self.phase == RunPhase.TRAIN:
+            self.decay_exploration()
+        epsilon = self.evaluation_epsilon if self.phase == RunPhase.TEST else self.epsilon
+
+        if self.discrete_controls:
+            top_action = np.argmax(action_values)
+            if np.random.rand() < epsilon:
+                return np.random.randint(self.action_space_size)
+            else:
+                return top_action
+        else:
+            noise = np.random.randn(1, self.action_space_size) * self.variance * self.action_abs_range
+            return np.squeeze(action_values + (np.random.rand() < epsilon) * noise)
+
+    def get_control_param(self):
+        return self.epsilon

exploration_policies/exploration_policy.py

@@ -0,0 +1,58 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import numpy as np
+from utils import *
+from configurations import *
+
+
+class ExplorationPolicy:
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        self.phase = RunPhase.HEATUP
+        self.action_space_size = tuning_parameters.env.action_space_size
+        self.action_abs_range = tuning_parameters.env_instance.action_space_abs_range
+        self.discrete_controls = tuning_parameters.env_instance.discrete_controls
+
+    def reset(self):
+        """
+        Used for resetting the exploration policy parameters when needed
+        :return: None
+        """
+        pass
+
+    def get_action(self, action_values):
+        """
+        Given a list of values corresponding to each action, 
+        choose one actions according to the exploration policy
+        :param action_values: A list of action values
+        :return: The chosen action
+        """
+        pass
+
+    def change_phase(self, phase):
+        """
+        Change between running phases of the algorithm
+        :param phase: Either Heatup or Train
+        :return: none
+        """
+        self.phase = phase
+
+    def get_control_param(self):
+        return 0

exploration_policies/greedy.py

@@ -0,0 +1,32 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+
+
+class Greedy(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+
+    def get_action(self, action_values):
+        return np.argmax(action_values)
+
+    def get_control_param(self):
+        return 0

exploration_policies/ou_process.py

@@ -0,0 +1,52 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import numpy as np
+from exploration_policies.exploration_policy import *
+
+# Based on on the description in:
+# https://math.stackexchange.com/questions/1287634/implementing-ornstein-uhlenbeck-in-matlab
+
+# Ornstein-Uhlenbeck process
+class OUProcess(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.action_space_size = tuning_parameters.env.action_space_size
+        self.mu = float(tuning_parameters.exploration.mu) * np.ones(self.action_space_size)
+        self.theta = tuning_parameters.exploration.theta
+        self.sigma = float(tuning_parameters.exploration.sigma) * np.ones(self.action_space_size)
+        self.state = np.zeros(self.action_space_size)
+        self.dt = tuning_parameters.exploration.dt
+
+    def reset(self):
+        self.state = np.zeros(self.action_space_size)
+
+    def noise(self):
+        x = self.state
+        dx = self.theta * (self.mu - x) * self.dt + self.sigma * np.random.randn(len(x)) * np.sqrt(self.dt)
+        self.state = x + dx
+        return self.state[0]
+
+    def get_action(self, action_values):
+        noise = self.noise()
+        return action_values.squeeze() + noise
+
+    def get_control_param(self):
+        return self.state[0]

exploration_policies/thompson_sampling.py

@@ -0,0 +1,35 @@
+#
+# Copyright (c) 2017 Intel Corporation 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from exploration_policies.exploration_policy import *
+
+
+class ThompsonSampling(ExplorationPolicy):
+    def __init__(self, tuning_parameters):
+        """
+        :param tuning_parameters: A Preset class instance with all the running paramaters
+        :type tuning_parameters: Preset
+        """
+        ExplorationPolicy.__init__(self, tuning_parameters)
+        self.action_space_size = tuning_parameters.env.action_space_size
+
+    def get_action(self, action_values):
+        q_values, values_uncertainty = action_values
+        sampled_q_values = np.random.normal(q_values, abs(values_uncertainty))
+        return np.argmax(sampled_q_values)
+
+    def get_control_param(self):
+        return 0