# Copyright 2020,2021 Sony Corporation.
# Copyright 2021 Sony Group 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 dataclasses import dataclass
from typing import Callable, Dict, Tuple
import numpy as np
from nnabla_rl.environment_explorer import EnvironmentExplorer, EnvironmentExplorerConfig
from nnabla_rl.environments.environment_info import EnvironmentInfo
def epsilon_greedy_action_selection(state, greedy_action_selector, random_action_selector, epsilon):
if np.random.rand() > epsilon:
# optimal action
return greedy_action_selector(state), True
else:
# random action
return random_action_selector(state), False
@dataclass
class NoDecayEpsilonGreedyExplorerConfig(EnvironmentExplorerConfig):
epsilon: float = 1.0
def __post_init__(self):
self._assert_between(self.epsilon, 0.0, 1.0, 'epsilon')
class NoDecayEpsilonGreedyExplorer(EnvironmentExplorer):
def __init__(self,
greedy_action_selector: Callable[[np.ndarray], Tuple[np.ndarray, Dict]],
random_action_selector: Callable[[np.ndarray], Tuple[np.ndarray, Dict]],
env_info: EnvironmentInfo,
config: NoDecayEpsilonGreedyExplorerConfig = NoDecayEpsilonGreedyExplorerConfig()):
super().__init__(env_info, config)
self._greedy_action_selector = greedy_action_selector
self._random_action_selector = random_action_selector
def action(self, step, state):
epsilon = self._config.epsilon
(action, info), _ = epsilon_greedy_action_selection(state,
self._greedy_action_selector,
self._random_action_selector,
epsilon)
return action, info
@dataclass
class LinearDecayEpsilonGreedyExplorerConfig(EnvironmentExplorerConfig):
"""
List of configurations for Linear decay epsilon-greedy explorer
Args:
initial_epsilon (float): Initial value of epsilon. Defaults to 1.0.
final_epsilon (float): Final value of epsilon after max_explore_steps.
This value must be smaller than initial_epsilon. Defaults to 0.05.
max_explore_steps (int): Number of steps to decay epsilon from initial_epsilon to final_epsilon.
Defaults to 1000000.
"""
initial_epsilon: float = 1.0
final_epsilon: float = 0.05
max_explore_steps: float = 1000000
def __post_init__(self):
self._assert_between(self.initial_epsilon, 0.0, 1.0, 'initial_epsilon')
self._assert_between(self.final_epsilon, 0.0, 1.0, 'final_epsilon')
self._assert_descending_order([self.initial_epsilon, self.final_epsilon], 'initial/final epsilon')
self._assert_positive(self.max_explore_steps, 'max_explore_steps')
[docs]class LinearDecayEpsilonGreedyExplorer(EnvironmentExplorer):
'''Linear decay epsilon-greedy explorer
Epsilon-greedy style explorer. Epsilon is linearly decayed until max_eplore_steps set in the config.
Args:
greedy_action_selector (Callable[[np.ndarray], Tuple[np.ndarray, Dict]]):
callable which computes greedy action with respect to current state.
random_action_selector (Callable[[np.ndarray], Tuple[np.ndarray, Dict]]):
callable which computes random action that can be executed in the environment.
env_info (:py:class:`EnvironmentInfo <nnabla_rl.environments.environment_info.EnvironmentInfo>`):
environment info
config (:py:class:`LinearDecayEpsilonGreedyExplorerConfig\
<nnabla_rl.environment_explorers.LinearDecayEpsilonGreedyExplorerConfig>`): the config of this class.
'''
def __init__(self,
greedy_action_selector: Callable[[np.ndarray], Tuple[np.ndarray, Dict]],
random_action_selector: Callable[[np.ndarray], Tuple[np.ndarray, Dict]],
env_info: EnvironmentInfo,
config: LinearDecayEpsilonGreedyExplorerConfig = LinearDecayEpsilonGreedyExplorerConfig()):
super().__init__(env_info, config)
self._greedy_action_selector = greedy_action_selector
self._random_action_selector = random_action_selector
[docs] def action(self, step, state):
epsilon = self._compute_epsilon(step)
(action, info), _ = epsilon_greedy_action_selection(state,
self._greedy_action_selector,
self._random_action_selector,
epsilon)
return action, info
def _compute_epsilon(self, step):
assert 0 <= step
delta_epsilon = step / self._config.max_explore_steps \
* (self._config.initial_epsilon - self._config.final_epsilon)
epsilon = self._config.initial_epsilon - delta_epsilon
return max(epsilon, self._config.final_epsilon)