A sample code used to solve BipedalWalker-v3 & BipedalWalkerHardcore-v3

• python 3.7
• pytorch 1.0.1 [ Warning ! ]
• (I ever used v1.7, then I waste a month to deal with it using the same code but without any network changes, please be careful!)
• gym 0.13.1

Agent uses the following hyperparameters:

gamma=0.99
batch_size=256
lr=5e-4
hidden_size=400
tau=0.005
alpha=0.2
reward_scale = 5 // reward *= reward_scale
capacity=2000000

1. Train from blank network

$ python main.py --train=1

2. Train from exist network

$ python main.py --train=1 --load=1

3. Play with the network

$ python main.py --load=1 --render=1

BiliBili: https://www.bilibili.com/video/BV1DK4y1j7Nz/

Pranjal Tandon (https://github.com/pranz24).