ray/rllib/tuned_examples/dqn/pong-dqn.yaml

29 lines
932 B
YAML

# You can expect ~20 reward within 1.1m timesteps / 2.1 hours on a K80 GPU
pong-deterministic-dqn:
env: PongDeterministic-v4
run: DQN
stop:
episode_reward_mean: 20
time_total_s: 7200
config:
# Works for both torch and tf.
framework: tf
num_gpus: 1
gamma: 0.99
lr: .0001
replay_buffer_config:
type: MultiAgentPrioritizedReplayBuffer
capacity: 50000
num_steps_sampled_before_learning_starts: 10000
rollout_fragment_length: 4
train_batch_size: 32
exploration_config:
epsilon_timesteps: 200000
final_epsilon: .01
model:
grayscale: True
zero_mean: False
dim: 42
# we should set compress_observations to True because few machines
# would be able to contain the replay buffers in memory otherwise
compress_observations: True