gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-18 19:50:17 +01:00
Code Issues Projects Releases Wiki Activity

rainbow dqn hyper-parameter updates

Browse Source
This commit is contained in:
Gal Leibovich
2018-08-30 20:41:33 +03:00
parent 07db625987
commit d862a3be83
2 changed files with 18 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
rl_coach/agents/rainbow_dqn_agent.py
View File

@@ -22,7 +22,8 @@ from rl_coach.agents.categorical_dqn_agent import CategoricalDQNAlgorithmParamet
CategoricalDQNAgent, CategoricalDQNAgentParameters CategoricalDQNAgent, CategoricalDQNAgentParameters
from rl_coach.agents.dqn_agent import DQNNetworkParameters from rl_coach.agents.dqn_agent import DQNNetworkParameters
from rl_coach.architectures.tensorflow_components.heads.rainbow_q_head import RainbowQHeadParameters from rl_coach.architectures.tensorflow_components.heads.rainbow_q_head import RainbowQHeadParameters
from rl_coach.architectures.tensorflow_components.middlewares.fc_middleware import FCMiddlewareParameters
from rl_coach.base_parameters import MiddlewareScheme
from rl_coach.exploration_policies.parameter_noise import ParameterNoiseParameters from rl_coach.exploration_policies.parameter_noise import ParameterNoiseParameters
from rl_coach.memories.non_episodic.prioritized_experience_replay import PrioritizedExperienceReplayParameters, \ from rl_coach.memories.non_episodic.prioritized_experience_replay import PrioritizedExperienceReplayParameters, \
PrioritizedExperienceReplay PrioritizedExperienceReplay
@@ -32,6 +33,7 @@ class RainbowDQNNetworkParameters(DQNNetworkParameters):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
self.heads_parameters = [RainbowQHeadParameters()] self.heads_parameters = [RainbowQHeadParameters()]
self.middleware_parameters = FCMiddlewareParameters(scheme=MiddlewareScheme.Empty)
class RainbowDQNAlgorithmParameters(CategoricalDQNAlgorithmParameters): class RainbowDQNAlgorithmParameters(CategoricalDQNAlgorithmParameters):
@@ -44,6 +46,11 @@ class RainbowDQNExplorationParameters(ParameterNoiseParameters):
super().__init__(agent_params) super().__init__(agent_params)
class RainbowDQNMemoryParameters(PrioritizedExperienceReplayParameters):
def __init__(self):
super().__init__()
class RainbowDQNAgentParameters(CategoricalDQNAgentParameters): class RainbowDQNAgentParameters(CategoricalDQNAgentParameters):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
@@ -58,8 +65,8 @@ class RainbowDQNAgentParameters(CategoricalDQNAgentParameters):
# Rainbow Deep Q Network - https://arxiv.org/abs/1710.02298 # Rainbow Deep Q Network - https://arxiv.org/abs/1710.02298
# Agent implementation is WIP. Currently has: # Agent implementation is WIP. Currently is composed of:
# 1. DQN # 1. NoisyNets
# 2. C51 # 2. C51
# 3. Prioritized ER # 3. Prioritized ER
# 4. DDQN # 4. DDQN

14
rl_coach/presets/Atari_Rainbow.py
View File

@@ -1,4 +1,3 @@
from rl_coach.agents.categorical_dqn_agent import CategoricalDQNAgentParameters
from rl_coach.agents.rainbow_dqn_agent import RainbowDQNAgentParameters from rl_coach.agents.rainbow_dqn_agent import RainbowDQNAgentParameters
from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters
from rl_coach.core_types import EnvironmentSteps, RunPhase from rl_coach.core_types import EnvironmentSteps, RunPhase
@@ -13,17 +12,20 @@ from rl_coach.schedules import LinearSchedule
#################### ####################
schedule_params = ScheduleParameters() schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentSteps(50000000) schedule_params.improve_steps = EnvironmentSteps(50000000)
schedule_params.steps_between_evaluation_periods = EnvironmentSteps(250000) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(1000000)
schedule_params.evaluation_steps = EnvironmentSteps(135000) schedule_params.evaluation_steps = EnvironmentSteps(125000)
schedule_params.heatup_steps = EnvironmentSteps(50000) schedule_params.heatup_steps = EnvironmentSteps(20000)
######### #########
# Agent # # Agent #
######### #########
agent_params = RainbowDQNAgentParameters() agent_params = RainbowDQNAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
agent_params.memory.beta = LinearSchedule(0.4, 1, 12500000) # 12.5M training iterations = 50M steps = 200M frames
agent_params.network_wrappers['main'].learning_rate = 0.0000625
agent_params.network_wrappers['main'].optimizer_epsilon = 1.5e-4
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(32000 // 4) # 32k frames
agent_params.memory.beta = LinearSchedule(0.4, 1, 12500000) # 12.5M training iterations = 50M steps = 200M frames
agent_params.memory.alpha = 0.5
############### ###############
# Environment # # Environment #