Benchmarks are being updated continuously.

Implementation and hyper-parameters of JORLDY can be different from official implementation

Atari

• Scores in the parentheses are score of the paper
• Red scores show best performance among the implemented algorithms for the environment
• Tensorboard smooth = 0.6
• Training info
  • DQN: training step = 10M, Paper training frame = 200M
  • PPO: training step = 10M, actors = 8, Paper training step = 10M, Paper actors = 8
  • Ape-X: training step = 10M, actors = 128, Paper environment frame = 22800M, Paper actors = 360

Mujoco

• Scores in the parentheses are score of the paper
• Red scores show best performance among the implemented algorithms for the environment
• Tensorboard smooth = 0.6
• Training info
  • DDPG: training step = 1M, Paper training frame = 2.5M
  • TD3: training step = 1M, Paper training step = 1M
  • SAC: training step = 1M, Paper training step = Different for each env (1M ~ 10M)
  • PPO: training step = 1M, actors = 32, Paper training step = 1M

Tensorborad

Tensorboard (score only)

Tensorboard (all metrics)