chore: import upstream snapshot with attribution

2026-07-13 13:17:40 +08:00
commit f1825c8ceb
10096 changed files with 2364182 additions and 0 deletions
@@ -0,0 +1,44 @@
+# @OldAPIStack
+# Given a SAC-generated offline file generated via:
+# rllib train -f examples/algorithms/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/env_runners/episode_return_mean: -700
+        timesteps_total: 800000
+    config:
+        # Works for both torch and tf.
+        framework: torch
+
+        # Set seed.
+        seed: 0
+
+        # Use one or more offline files or "input: sampler" for online learning.
+        input: 'dataset'
+        input_config:
+            paths: ["offline/tests/data/pendulum/enormous.zip"]
+            format: 'json'
+        # 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
+        bc_iters: 100
+        num_env_runners: 2
+        min_time_s_per_iteration: 10
+
+        metrics_num_episodes_for_smoothing: 5
+
+        # Evaluate in an actual environment.
+        evaluation_interval: 1
+        evaluation_num_env_runners: 2
+        evaluation_duration: 10
+        evaluation_parallel_to_training: true
+        evaluation_config:
+            input: sampler
+            explore: False