BipedalWalker-V3 with TD3 actor critic (MLP nets)

Motivation

• The main goal of this project is that to solve the problem characterized by the environment BipedalWalker-v3. ( https://gym.openai.com/envs/BipedalWalker-v2/ ).

Algorithms used to train the Agent

• TD3

Dependencies

• Python 3.8
• gym 0.17.2
• box2d 2.3.10
• torch 1.5.0
• numpy 1.18.4
• Matplotlib 3.2.1
• Pillow 7.1.2

Refernces

• “ Addressing Function Approximation Error in Actor-Critic Methods ” https://arxiv.org/pdf/1802.09477.pdf
• https://towardsdatascience.com/td3-learning-to-run-with-ai-40dfc512f93
• https://gym.openai.com/envs/BipedalWalker-v2/

Model Training

• 100 itr : https://drive.google.com/file/d/1FmlKrmKwmio7BcBCvZfEW8P6IN09jkEN/view?usp=sharing
• 500+ itr : https://drive.google.com/file/d/17FVSPo-tvA1g9XPOdRxU_b8jjomagbk2/view?usp=sharing