gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 11:10:20 +01:00
Code Issues Projects Releases Wiki Activity

preventing the evaluation agent from getting stuck in bad policies by updating from the global network during episodes

Browse Source
This commit is contained in:
Itai Caspi
2017-10-25 10:28:45 +03:00
parent d3c6860421
commit 39cf78074c
2 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
agents/agent.py
View File

@@ -457,6 +457,11 @@ class Agent:
while not episode_ended: while not episode_ended:
episode_ended = self.act(phase=RunPhase.TEST) episode_ended = self.act(phase=RunPhase.TEST)
if keep_networks_synced \
and self.total_steps_counter % self.tp.agent.update_evaluation_agent_network_after_every_num_steps:
for network in self.networks:
network.sync()
if self.tp.visualization.dump_gifs and self.total_reward_in_current_episode > max_reward_achieved: if self.tp.visualization.dump_gifs and self.total_reward_in_current_episode > max_reward_achieved:
max_reward_achieved = self.total_reward_in_current_episode max_reward_achieved = self.total_reward_in_current_episode
frame_skipping = int(5/self.tp.env.frame_skip) frame_skipping = int(5/self.tp.env.frame_skip)

1
configurations.py
View File

@@ -72,6 +72,7 @@ class AgentParameters:
# Agent parameters # Agent parameters
num_consecutive_playing_steps = 1 num_consecutive_playing_steps = 1
num_consecutive_training_steps = 1 num_consecutive_training_steps = 1
update_evaluation_agent_network_after_every_num_steps = 3000
bootstrap_total_return_from_old_policy = False bootstrap_total_return_from_old_policy = False
n_step = -1 n_step = -1
num_episodes_in_experience_replay = 200 num_episodes_in_experience_replay = 200