mirror of
https://github.com/gryf/coach.git
synced 2025-12-17 19:20:19 +01:00
118 lines
6.1 KiB
Python
118 lines
6.1 KiB
Python
#
|
|
# 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 typing import List, Union, Tuple
|
|
|
|
from rl_coach.base_parameters import AgentParameters, VisualizationParameters, TaskParameters, \
|
|
PresetValidationParameters
|
|
from rl_coach.core_types import EnvironmentSteps
|
|
from rl_coach.environments.environment import EnvironmentParameters, Environment
|
|
from rl_coach.graph_managers.graph_manager import GraphManager, ScheduleParameters
|
|
from rl_coach.level_manager import LevelManager
|
|
from rl_coach.utils import short_dynamic_import
|
|
|
|
|
|
class HRLGraphManager(GraphManager):
|
|
"""
|
|
A simple HRL graph manager creates a deep hierarchy with a single composite agent per hierarchy level, and a single
|
|
environment which is interacted with.
|
|
"""
|
|
def __init__(self, agents_params: List[AgentParameters], env_params: EnvironmentParameters,
|
|
schedule_params: ScheduleParameters, vis_params: VisualizationParameters,
|
|
consecutive_steps_to_run_each_level: Union[EnvironmentSteps, List[EnvironmentSteps]],
|
|
preset_validation_params: PresetValidationParameters = PresetValidationParameters()):
|
|
"""
|
|
:param agents_params: the parameters of all the agents in the hierarchy starting from the top level of the
|
|
hierarchy to the bottom level
|
|
:param env_params: the parameters of the environment
|
|
:param schedule_params: the parameters for scheduling the graph
|
|
:param vis_params: the visualization parameters
|
|
:param consecutive_steps_to_run_each_level: the number of time steps that each level is ran.
|
|
for example, when the top level gives the bottom level a goal, the bottom level can act for
|
|
consecutive_steps_to_run_each_level steps and try to reach that goal. This is expected to be either
|
|
an EnvironmentSteps which will be used for all levels, or an EnvironmentSteps for each level as a list.
|
|
"""
|
|
super().__init__('hrl_graph', schedule_params, vis_params)
|
|
self.agents_params = agents_params
|
|
self.env_params = env_params
|
|
self.preset_validation_params = preset_validation_params
|
|
if isinstance(consecutive_steps_to_run_each_level, list):
|
|
if len(consecutive_steps_to_run_each_level) != len(self.agents_params):
|
|
raise ValueError("If the consecutive_steps_to_run_each_level is given as a list, it should match "
|
|
"the number of levels in the hierarchy. Alternatively, it is possible to use a single "
|
|
"value for all the levels, by passing an EnvironmentSteps")
|
|
elif isinstance(consecutive_steps_to_run_each_level, EnvironmentSteps):
|
|
self.consecutive_steps_to_run_each_level = [consecutive_steps_to_run_each_level] * len(self.agents_params)
|
|
|
|
for agent_params in agents_params:
|
|
agent_params.visualization = self.visualization_parameters
|
|
if agent_params.input_filter is None:
|
|
agent_params.input_filter = self.env_params.default_input_filter()
|
|
if agent_params.output_filter is None:
|
|
agent_params.output_filter = self.env_params.default_output_filter()
|
|
|
|
if len(self.agents_params) < 2:
|
|
raise ValueError("The HRL graph manager must receive the agent parameters for at least two levels of the "
|
|
"hierarchy. Otherwise, use the basic RL graph manager.")
|
|
|
|
def _create_graph(self, task_parameters: TaskParameters) -> Tuple[List[LevelManager], List[Environment]]:
|
|
self.env_params.seed = task_parameters.seed
|
|
env = short_dynamic_import(self.env_params.path)(**self.env_params.__dict__,
|
|
visualization_parameters=self.visualization_parameters)
|
|
|
|
for agent_params in self.agents_params:
|
|
agent_params.task_parameters = task_parameters
|
|
|
|
# we need to build the hierarchy in reverse order (from the bottom up) in order for the spaces of each level
|
|
# to be known
|
|
level_managers = []
|
|
current_env = env
|
|
# out_action_space = env.action_space
|
|
for level_idx, agent_params in reversed(list(enumerate(self.agents_params))):
|
|
# TODO: the code below is specific for HRL on observation scale
|
|
# in action space
|
|
# if level_idx == 0:
|
|
# # top level agents do not get directives
|
|
# in_action_space = None
|
|
# else:
|
|
# pass
|
|
|
|
# attention_size = (env.state_space['observation'].shape - 1)//4
|
|
# in_action_space = AttentionActionSpace(shape=2, low=0, high=env.state_space['observation'].shape - 1,
|
|
# forced_attention_size=attention_size)
|
|
# agent_params.output_filter.action_filters['masking'].set_masking(0, attention_size)
|
|
|
|
agent_params.name = "agent_{}".format(level_idx)
|
|
agent_params.is_a_highest_level_agent = level_idx == 0
|
|
agent = short_dynamic_import(agent_params.path)(agent_params)
|
|
|
|
level_manager = LevelManager(
|
|
agents=agent,
|
|
environment=current_env,
|
|
real_environment=env,
|
|
steps_limit=self.consecutive_steps_to_run_each_level[level_idx],
|
|
should_reset_agent_state_after_time_limit_passes=level_idx > 0,
|
|
name="level_{}".format(level_idx)
|
|
)
|
|
current_env = level_manager
|
|
level_managers.insert(0, level_manager)
|
|
|
|
# out_action_space = in_action_space
|
|
|
|
return level_managers, [env]
|
|
|
|
|