Skip to content

Pytorch implementation of intrinsic curiosity module with proximal policy optimization

Notifications You must be signed in to change notification settings

chagmgang/pytorch_ppo_rl

Repository files navigation

This Repository is Reinforcement Learning related with PPO

This Repository is Reinforcece Learning Implementation related with PPO. The framework used in this Repository is Pytorch. The multi-processing method is basically built in. The agents are trained by PAAC(Parallel Advantage Actor Critic) strategy.

1. Multi-processing MLP Proximal Policy Optimization

  • Script : LunarLander_ppo.py
  • Environment : LunarLander-v2
  • Orange : 8 Process, Blue : 4 Process, Red : 1 Process
LunarLander-v2

2. Multi-processing CNN Proximal Policy Opimization

  • Script : Breakout_ppo.py
  • Environment : BreakoutDeterministic-v4
  • Red: 8 Process, Blue: 4 Process, Orange: 1 Process
BreakoutDeterministic-v4

3. Multi-processing CNN Proximal Policy Opitimization with Intrinsic Curiosity Module

  • Script : Breakout_ppo_icm.py
  • Environment : BreakoutNoFrameskip-v4(handled by custom environment)
  • With no environment Reward
  • Because the game initial key is not selected, the peak point and performance drop is generated.
  • Left : Comparison between (extrinsic reward and intrinsic, oragne) and (only intrinsic reward, gray), the average of three times of experiment
  • Right : only intrinsic reward
  • 32 process
BreakoutNoFrameskip-v4(handled by custom environment)

4. Multi-processing Mlp Proximal Policy Opitimization with Intrinsic Curiosity Module

  • Script : MountainCar_ppo_icm.py
  • Environment : MountainCart-v0
  • With no environment Reward
  • 32 process
MountainCart-v0

5. Unity MLAgents Mlp Proximal Policy Optimization with Intrinsic Curiosity Module

  • Script : PushBlock_ppo_icm.py
  • Environment : PushBlock
  • 32 Environment, PAAC
  • orange : 0.5int + 0.5ext, blue : only int, Red : only ext
  • reward shaping for sparse-reward environment : sucess - 1, others - 0
  • The environment has not sparsed-reward property even if the reward is engineered to two categories(0, 1)
PushBlock

6. Unity MLAgents Mlp Proximal Policy Optimization with Intrinsic Curiosity Module

  • Script : Pyramid_ppo_icm.py
  • Environment : Pyramid
  • 16 Environment, PAAC
  • orange : only ext, blue : 0.01int + 0.99ext
Pyramid

Reference

[1] mario_rl

[2] Proximal Policy Optimization

[2] Efficient Parallel Methods for Deep Reinforcement Learning

[3] High-Dimensional Continuous Control Using Generalized Advantage Estimation

[4] Curiosity-driven Exploration by Self-supervised Prediction

[5] Large-Scale Study of Curiosity-Driven Learning

[6] curiosity-driven-exploration-pytorch

[7] ml-agents

[8] Unity: A General Platform for Intelligent Agents

[9] Solving sparse-reward tasks with Curiosity

Releases

No releases published

Packages

No packages published

Languages