Experiment1

README.md

@@ -41,12 +41,12 @@ agent.load(path)
 ```python
 # https://skrl.readthedocs.io/en/latest/modules/skrl.agents.ppo.html#configuration-and-hyperparameters
 cfg_ppo = PPO_DEFAULT_CONFIG.copy()
-cfg_ppo["rollouts"] = 16  # memory_size
-cfg_ppo["learning_epochs"] = 5
-cfg_ppo["mini_batches"] = 4  # 16 * 8192 / 32768
+cfg_ppo["rollouts"] = 20 # memory_size
+cfg_ppo["learning_epochs"] = 10
+cfg_ppo["mini_batches"] = 4 # 16 * 8192 / 32768
 cfg_ppo["discount_factor"] = 0.99
-cfg_ppo["lambda"] = 0.95
-cfg_ppo["learning_rate"] = 5e-4
+cfg_ppo["lambda"] = 1
+cfg_ppo["learning_rate"] = 6e-4
 cfg_ppo["learning_rate_scheduler"] = KLAdaptiveRL
 cfg_ppo["learning_rate_scheduler_kwargs"] = {"kl_threshold": 0.016}
 cfg_ppo["random_timesteps"] = 0
@@ -56,7 +56,7 @@ cfg_ppo["ratio_clip"] = 0.2
 cfg_ppo["value_clip"] = 0.2
 cfg_ppo["clip_predicted_values"] = True
 cfg_ppo["entropy_loss_scale"] = 0.0
-cfg_ppo["value_loss_scale"] = 2.0
+cfg_ppo["value_loss_scale"] = 2.5
 cfg_ppo["kl_threshold"] = 0
 cfg_ppo["rewards_shaper"] = lambda rewards, timestep, timesteps: rewards * 0.01
 cfg_ppo["state_preprocessor"] = RunningStandardScaler
@@ -66,4 +66,5 @@ cfg_ppo["value_preprocessor_kwargs"] = {"size": 1, "device": device}
 # logging to TensorBoard and writing checkpoints
 cfg_ppo["experiment"]["write_interval"] = 800
 cfg_ppo["experiment"]["checkpoint_interval"] = 8000
+
 ```