From bc120730e58cf933e2b4ca0837a7cda2a38c65a1 Mon Sep 17 00:00:00 2001 From: Sven Mika Date: Thu, 12 Mar 2020 04:31:27 +0100 Subject: [PATCH] [RLlib] PPO(torch) on CartPole not tuned well enough for consistent learning (#7556) --- .../regression_tests/cartpole-ppo-tf.yaml | 9 ++++++++- .../regression_tests/cartpole-ppo-torch.yaml | 15 +++++++++++---- 2 files changed, 19 insertions(+), 5 deletions(-) diff --git a/rllib/tuned_examples/regression_tests/cartpole-ppo-tf.yaml b/rllib/tuned_examples/regression_tests/cartpole-ppo-tf.yaml index 09bf5dafd..315b7dd44 100644 --- a/rllib/tuned_examples/regression_tests/cartpole-ppo-tf.yaml +++ b/rllib/tuned_examples/regression_tests/cartpole-ppo-tf.yaml @@ -5,6 +5,13 @@ cartpole-ppo-tf: episode_reward_mean: 150 timesteps_total: 100000 config: + gamma: 0.99 + lr: 0.0003 num_workers: 1 - batch_mode: complete_episodes observation_filter: MeanStdFilter + num_sgd_iter: 6 + vf_share_layers: true + vf_loss_coeff: 0.01 + model: + fcnet_hiddens: [32] + fcnet_activation: linear diff --git a/rllib/tuned_examples/regression_tests/cartpole-ppo-torch.yaml b/rllib/tuned_examples/regression_tests/cartpole-ppo-torch.yaml index 3d3ad4291..ec953a653 100644 --- a/rllib/tuned_examples/regression_tests/cartpole-ppo-torch.yaml +++ b/rllib/tuned_examples/regression_tests/cartpole-ppo-torch.yaml @@ -1,11 +1,18 @@ -cartpole-ppo-tf: +cartpole-ppo-torch: env: CartPole-v0 run: PPO stop: episode_reward_mean: 150 timesteps_total: 100000 config: - num_workers: 1 - batch_mode: complete_episodes - observation_filter: MeanStdFilter use_pytorch: true + gamma: 0.99 + lr: 0.0003 + num_workers: 1 + observation_filter: MeanStdFilter + num_sgd_iter: 6 + vf_share_layers: true + vf_loss_coeff: 0.01 + model: + fcnet_hiddens: [32] + fcnet_activation: linear