Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 260M agent steps
To reproduce run:
python main.py Boxing-v0 \
--initial_lr 0.00025 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 160M agent steps
To reproduce run:
python main.py Boxing-v0 \
--initial_lr 0.00025 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using DQN+CTS implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 18M agent steps using epsilon of .01
To reproduce run:
python main.py Freeway-v0 \
--load_config config/dqn-cts.yaml \
--epsilon_annealing_steps 200000 \
--clip_norm_type ignore \
--q_target_update_steps 20000 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 300M agent steps
To reproduce run:
python main.py WizardOfWor-v0 \
--initial_lr 0.0007 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 80M agent steps
To reproduce run:
python main.py Boxing-v0 \
--initial_lr 0.00025 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 56M agent steps
To reproduce run:
python main.py Boxing-v0 \
--initial_lr 0.00025 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 300M agent steps
To reproduce run:
python main.py YarsRevenge-v0 \
--initial_lr 0.0007 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 180M agent steps
To reproduce run:
python main.py YarsRevenge-v0 \
--initial_lr 0.0007 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Produced using PGQ implementation from tensorflow-rl at commit daf75a33a20aae461a63c5b650b61216117b3f7b. Evaluation generated from checkpoint at 80M agent steps
To reproduce run:
python main.py YarsRevenge-v0 \
--initial_lr 0.0007 \
--momentum .99 \
--clip_norm_type ignore \
--frame_skip 2 4 \
Generated using a3c-lstm implementation from https://github.com/steveKapturowski/tensorflow-rl Trained for 75M agent steps with 16 agents and --max_local_steps=20 --td_lambda=.97