jaxgmg2_3phase_optim_state

jaxgmg2_3phase_optim_state

32 RL agent checkpoints trained on the JaxGMG maze environment with discount_rate=0.98, with full optimizer state saved at each checkpoint. Primary runs for the 3-phase training regime analysis. Similar to jaxgmg2_3phase_unique but with optimizer state logging enabled.

• 16 runs with alpha=0.6 (run_id 15-30)
• 16 runs with alpha=1.0 (run_id 15-30)

WandB: https://wandb.ai/devinterp/jaxgmg2_3phase_optim_state

Sweep

run_id sweep: 15-30 for each alpha value. Seed is derived from run_id via: seed = int(discount_rate*100)*10000 + int(alpha*10)*100 + run_id e.g. alpha=0.6, run_id=15 -> seed=980615; alpha=1.0, run_id=15 -> seed=981015.

Shared Hyperparams

rl_action=train
alpha=0.6 or 1.0
discount_rate=0.98
lr=5e-05
num_total_env_steps=10000000000
num_rollout_steps=64
num_levels=9600
cheese_loc=any
env_layout=open
env_size=13
mask_type=first_episode
use_prev_action=False
grad_acc_per_chunk=5
log_optimizer_state=True
eval_schedule=0:1,250:2,500:5,2000:10
seed_formula={int(discount_rate*100):02d}{int(alpha*10):02d}{run_id:02d}
f_str_ckpt=al_{alpha}_g_{discount_rate}_id_{run_id}_seed_{seed}
ckpt_dir=jaxgmg2_3phase_optim_state
wandb_project=jaxgmg2_3phase_optim_state
use_wandb=True
use_hf=True

Naming Schema

Checkpoints are named al_{alpha}_g_0.98_id_{run_id}_seed_{seed}.

Reproduced with

See train.yaml in this repository. Run with:

make run projects/rl/experiments/al_0.6_g_0.98/jobs/train_optim_state.yaml

from the timaeus monorepo.