Ubastic/Bootstrapped-Meta-Learning

Replication of Bootstrapped Meta-Learning

Bootstrapped Meta-Learning Replication

Display

• A2C Baseline
  • 
  • 
• Meta-Gradient Reinforcement Learning
  • 
  • 
  • 
• Bootstrapped Meta-Learning
  • 
  • 
  • 

How To Use

A2C

$ python main.py --algo A2C --num-seeds 1 --epsilon-EN 3e-1 --lr 1e-1 --gamma 0.99 --log-dir "logs" --max-steps 6400000

Both SGD and Adam are supported, however, it is not encouraged to use Adam otherwise you won't see desired result.

MGRL

$ python main.py --algo MGRL --num-seeds 1 --lr 1e-1 --meta-lr 1e-4 -epsilon-meta 0.12 --gamma 0.99 --T 15 --max-steps 6400000

BMG

$ python main.py --algo BMG -num-seeds 1 --lr 1e-1 --meta-lr 1e-4 --gamma 0.99 --K 7 --L 9 --max-steps 6400000

We use 7 and 9 corresponding to K and L for default, for details, you can redirect to the original paper Bootstrapped Meta-Learning

Requirements

$ pip install requirements.txt