index

rl-clarity model weights

For interfaces, see here.

Each model <relpath>/<name> has 5 associated files:

<relpath>/<name>.jd: checkpoint file from which the other files are derived using this library
<relpath>/rl-clarity/<name>.model.pb: Lucid model save file
<relpath>/rl-clarity/<name>.metadata.jd: joblib file containing metadata
<relpath>/rl-clarity/<name>.observations.jd: joblib file containing observations sampled infrequently from the agent playing the game, intended to be used for NMF
<relpath>/rl-clarity/<name>.trajectories.jd: joblib file containing observations, actions, rewards and first-in-episode booleans from contiguous trajectories of the agent playing the game

All paths are relative to directory of the current page (ending in models/).

Files can be loaded directly from their paths using lucid.scratch.rl_util.load_joblib for joblib files and lucid.modelzoo.vision_base.Model.load for Lucid model save files.

Original CoinRun model:

coinrun

Finite numbers of training levels:

Run 1
- 100 levels: finite_levels/run_1/coinrun_100
- 300 levels: finite_levels/run_1/coinrun_300
- 1,000 levels: finite_levels/run_1/coinrun_1000
- 3,000 levels: finite_levels/run_1/coinrun_3000
- 10,000 levels: finite_levels/run_1/coinrun_10000
- 30,000 levels: finite_levels/run_1/coinrun_30000
- 100,000 levels: finite_levels/run_1/coinrun_100000

Run 2
- 100 levels: finite_levels/run_2/coinrun_100
- 300 levels: finite_levels/run_2/coinrun_300
- 1,000 levels: finite_levels/run_2/coinrun_1000
- 3,000 levels: finite_levels/run_2/coinrun_3000
- 10,000 levels: finite_levels/run_2/coinrun_10000
- 30,000 levels: finite_levels/run_2/coinrun_30000
- 100,000 levels: finite_levels/run_2/coinrun_100000

Edited models:

Saw obstacle edit: edit/coinrun_saw_edit
Enemy moving left edit: edit/coinrun_enemy_edit

Procgen Benchmark models:

CoinRun: procgen/coinrun
StarPilot: procgen/starpilot
CaveFlyer: procgen/caveflyer
Dodgeball: procgen/dodgeball
FruitBot: procgen/fruitbot
Chaser: procgen/chaser
Miner: procgen/miner
Jumper: procgen/jumper
Leaper: procgen/leaper
Maze: procgen/maze
BigFish: procgen/bigfish
Heist: procgen/heist
Climber: procgen/climber
Plunder: procgen/plunder
Ninja: procgen/ninja
BossFight: procgen/bossfight

IMPALA architecture modifications:

Residual, pool size 3, max pooling (original): impala/coinrun_ablation_0
Non-residual, pool size 3, max pooling: impala/coinrun_ablation_1
Residual, pool size 2, max pooling: impala/coinrun_ablation_2
Non-residual, pool size 2, max pooling: impala/coinrun_ablation_3
Residual, pool size 3, L2 pooling: impala/coinrun_ablation_4
Non-residual, pool size 3, L2 pooling: impala/coinrun_ablation_5
Residual, pool size 2, L2 pooling: impala/coinrun_ablation_6
Non-residual, pool size 2, L2 pooling: impala/coinrun_ablation_7

Other:

Procedurally-generated CoinRun assets: coinrun_procgen