gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 19:20:19 +01:00
Code Issues Projects Releases Wiki Activity

bug fix - fixed C51 presets hyperparameters

Browse Source
This commit is contained in:
Itai Caspi
2017-11-10 13:22:00 +02:00
parent 3c330768f0
commit 8d9ee4ea2b
1 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
presets.py
View File

@@ -70,7 +70,6 @@ class Doom_Basic_DQN(Preset):
self.num_heatup_steps = 1000 self.num_heatup_steps = 1000
class Doom_Basic_QRDQN(Preset): class Doom_Basic_QRDQN(Preset):
def __init__(self): def __init__(self):
Preset.__init__(self, QuantileRegressionDQN, Doom, ExplorationParameters) Preset.__init__(self, QuantileRegressionDQN, Doom, ExplorationParameters)
@@ -362,6 +361,9 @@ class CartPole_C51(Preset):
self.num_heatup_steps = 1000 self.num_heatup_steps = 1000
self.exploration.epsilon_decay_steps = 3000 self.exploration.epsilon_decay_steps = 3000
self.agent.discount = 1.0 self.agent.discount = 1.0
# self.env.reward_scaling = 20.
self.agent.v_min = 0.0
self.agent.v_max = 200.0
self.test = True self.test = True
self.test_max_step_threshold = 150 self.test_max_step_threshold = 150
@@ -408,14 +410,16 @@ class Breakout_C51(Preset):
self.agent.num_steps_between_copying_online_weights_to_target = 10000 self.agent.num_steps_between_copying_online_weights_to_target = 10000
self.learning_rate = 0.00025 self.learning_rate = 0.00025
self.agent.num_transitions_in_experience_replay = 1000000 self.agent.num_transitions_in_experience_replay = 1000000
self.exploration.initial_epsilon = 0.01 self.exploration.initial_epsilon = 1.0
self.exploration.final_epsilon = 0.01 self.exploration.final_epsilon = 0.01
self.exploration.epsilon_decay_steps = 1000000 self.exploration.epsilon_decay_steps = 1000000
self.env.reward_clipping_max = 1.0
self.env.reward_clipping_min = -1.0
self.exploration.evaluation_policy = 'EGreedy' self.exploration.evaluation_policy = 'EGreedy'
self.exploration.evaluation_epsilon = 0.001 self.exploration.evaluation_epsilon = 0.001
self.num_heatup_steps = 50000 self.num_heatup_steps = 50000
self.evaluation_episodes = 25 self.evaluation_episodes = 1
self.evaluate_every_x_episodes = 10000 self.evaluate_every_x_episodes = 5000000
class Atari_DQN_TestBench(Preset): class Atari_DQN_TestBench(Preset):