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

new feature - implementation of Quantile Regression DQN (https://arxiv.org/pdf/1710.10044v1.pdf)

Browse Source 
API change - Distributional DQN renamed to Categorical DQN
This commit is contained in:
Itai Caspi
2017-11-01 15:09:07 +02:00
parent 1ad6262307
commit a8bce9828c
10 changed files with 157 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
presets.py
View File

@@ -70,6 +70,18 @@ class Doom_Basic_DQN(Preset):
self.num_heatup_steps = 1000
class Doom_Basic_QRDQN(Preset):
def __init__(self):
Preset.__init__(self, QuantileRegressionDQN, Doom, ExplorationParameters)
self.env.level = 'basic'
self.agent.num_steps_between_copying_online_weights_to_target = 1000
self.learning_rate = 0.00025
self.agent.num_episodes_in_experience_replay = 200
self.num_heatup_steps = 1000
class Doom_Basic_OneStepQ(Preset):
def __init__(self):
Preset.__init__(self, NStepQ, Doom, ExplorationParameters)
@@ -340,9 +352,9 @@ class CartPole_DQN(Preset):
self.test_min_return_threshold = 150
class CartPole_DistributionalDQN(Preset):
class CartPole_C51(Preset):
def __init__(self):
Preset.__init__(self, DistributionalDQN, GymVectorObservation, ExplorationParameters)
Preset.__init__(self, CategoricalDQN, GymVectorObservation, ExplorationParameters)
self.env.level = 'CartPole-v0'
self.agent.num_steps_between_copying_online_weights_to_target = 100
self.learning_rate = 0.00025
@@ -356,6 +368,18 @@ class CartPole_DistributionalDQN(Preset):
self.test_min_return_threshold = 150
class CartPole_QRDQN(Preset):
def __init__(self):
Preset.__init__(self, QuantileRegressionDQN, GymVectorObservation, ExplorationParameters)
self.env.level = 'CartPole-v0'
self.agent.num_steps_between_copying_online_weights_to_target = 100
self.learning_rate = 0.00025
self.agent.num_episodes_in_experience_replay = 200
self.num_heatup_steps = 1000
self.exploration.epsilon_decay_steps = 3000
self.agent.discount = 1.0
# The below preset matches the hyper-parameters setting as in the original DQN paper.
# This a very resource intensive preset, and might easily blow up your RAM (> 100GB of usage).
# Try reducing the number of transitions in the experience replay (50e3 might be a reasonable number to start with),
@@ -377,9 +401,9 @@ class Breakout_DQN(Preset):
self.evaluate_every_x_episodes = 100
class Breakout_DistributionalDQN(Preset):
class Breakout_C51(Preset):
def __init__(self):
Preset.__init__(self, DistributionalDQN, Atari, ExplorationParameters)
Preset.__init__(self, CategoricalDQN, Atari, ExplorationParameters)
self.env.level = 'BreakoutDeterministic-v4'
self.agent.num_steps_between_copying_online_weights_to_target = 10000
self.learning_rate = 0.00025