mirror of
https://github.com/vale981/ray
synced 2025-03-12 14:16:39 -04:00
53 lines
1.8 KiB
YAML
53 lines
1.8 KiB
YAML
long-term-satisfaction-recsim-env-slateq:
|
|
env: ray.rllib.examples.env.recommender_system_envs_with_recsim.LongTermSatisfactionRecSimEnv
|
|
run: SlateQ
|
|
stop:
|
|
# Random baseline rewards:
|
|
# num_candidates=20; slate_size=2; resample=true: ~951
|
|
# num_candidates=50; slate_size=3; resample=true: ~946
|
|
evaluation/episode_reward_mean: 1000.0
|
|
timesteps_total: 200000
|
|
config:
|
|
# SlateQ only supported for torch so far.
|
|
framework: torch
|
|
|
|
metrics_num_episodes_for_smoothing: 1000
|
|
|
|
# RLlib/RecSim wrapper specific settings:
|
|
env_config:
|
|
config:
|
|
# Each step, sample `num_candidates` documents using the env-internal
|
|
# document sampler model (a logic that creates n documents to select
|
|
# the slate from).
|
|
resample_documents: true
|
|
num_candidates: 50
|
|
# How many documents to recommend (out of `num_candidates`) each
|
|
# timestep?
|
|
slate_size: 2
|
|
# Should the action space be purely Discrete? Useful for algos that
|
|
# don't support MultiDiscrete (e.g. DQN or Bandits).
|
|
# SlateQ handles MultiDiscrete action spaces.
|
|
convert_to_discrete_action_space: false
|
|
seed: 42
|
|
|
|
exploration_config:
|
|
type: SlateSoftQ
|
|
temperature: 0.7
|
|
|
|
hiddens: [256, 256]
|
|
|
|
num_workers: 0
|
|
num_gpus: 0
|
|
|
|
lr_choice_model: 0.002
|
|
lr_q_model: 0.001
|
|
|
|
target_network_update_freq: 800
|
|
tau: 1.0
|
|
|
|
# Evaluation settings.
|
|
evaluation_interval: 1
|
|
evaluation_num_workers: 4
|
|
evaluation_duration: 200
|
|
evaluation_duration_unit: episodes
|
|
evaluation_parallel_to_training: true
|