ray/rllib/tuned_examples/pg/cartpole-crashing-with_remote-envs-pg.yaml

29 lines
1.2 KiB
YAML

cartpole-crashing-with-remote-envs-pg:
env: ray.rllib.examples.env.cartpole_crashing.CartPoleCrashing
run: PG
stop:
episode_reward_mean: 35.0
timesteps_total: 25000
config:
# Works for both torch and tf.
framework: tf
env_config:
config:
p_crash: 0.0
# Crash all envs always exactly after n steps.
crash_after_n_steps: 60
# Time for the env to initialize when newly created.
# Every time a remote sub-environment crashes, a new env is created
# instead that will take this long (sleep) to "initialize".
init_time_s: 2.0
num_workers: 4
num_envs_per_worker: 3
rollout_fragment_length: 50
#train_batch_size: 32
# Use parallel remote envs.
remote_worker_envs: true
# Switch on resiliency for failed sub environments (within a vectorized stack).
restart_failed_sub_environments: true
# Also ignore worker failures, however, this should not matter as we don't
# expect any due to catching individual sub-env failures.
ignore_worker_failures: true