mirror of
https://github.com/vale981/ray
synced 2025-03-06 10:31:39 -05:00
e7f9e8ceec
* Fix trainer timestep reporting for offline agents like CQL. * wip. * extend timesteps_total to 200K for learning_tests_pendulum_cql test Co-authored-by: sven1977 <svenmika1977@gmail.com>
36 lines
1.1 KiB
YAML
36 lines
1.1 KiB
YAML
# Given a SAC-generated offline file generated via:
|
|
# rllib train -f tuned_examples/sac/pendulum-sac.yaml --no-ray-ui
|
|
|
|
# Pendulum CQL can attain ~ -300 reward in 10k from that file.
|
|
pendulum-cql:
|
|
env: Pendulum-v1
|
|
run: CQL
|
|
stop:
|
|
evaluation/episode_reward_mean: -700
|
|
timesteps_total: 200000
|
|
config:
|
|
# Works for both torch and tf.
|
|
framework: tf
|
|
|
|
# Use one or more offline files or "input: sampler" for online learning.
|
|
input: ["tests/data/pendulum/enormous.zip"]
|
|
# Our input file above comes from an SAC run. Actions in there
|
|
# are already normalized (produced by SquashedGaussian).
|
|
actions_in_input_normalized: true
|
|
clip_actions: true
|
|
|
|
twin_q: true
|
|
train_batch_size: 2000
|
|
learning_starts: 0
|
|
bc_iters: 100
|
|
|
|
metrics_smoothing_episodes: 5
|
|
|
|
# Evaluate in an actual environment.
|
|
evaluation_interval: 1
|
|
evaluation_num_workers: 2
|
|
evaluation_num_episodes: 10
|
|
evaluation_parallel_to_training: true
|
|
evaluation_config:
|
|
input: sampler
|
|
explore: False
|