Meta_RL_for_SAS:

This project is using Meta Reinforcement learning to enhance the adaptability of self-learning adaptive system (SLAS).

The process of optimizing and the adaptation process from meta parameters to optimal parameters:

(This is adapted from MAML paper.)

To avoid any conflict with your existing Python setup, and to keep this project self-contained, it is suggested to work in a virtual environment with virtualenv. To install virtualenv:

pip install --upgrade virtualenv

Create a virtual environment, activate it and install the requirements in requirements.txt.

virtualenv venv
source venv/bin/activate
pip install -r requirements.txt

• Python 3.5 or above
• PyTorch 1.3
• Gym 0.15

You can use the train.py script in order to run reinforcement learning experiments with MAML. Note that by default, logs are available in train.py but are not saved (eg. the returns during meta-training). For example, to run the script on mdp-complex:

python train.py --config configs/maml/mdp/mdp-complex.yaml --output-folder mdp-complex/ --seed 2 --num-workers 4

Once you have meta-trained the policy, you can test it on the same environment using test.py:

python test-my-new.py --config mdp-complex/config.json --policy mdp-complex/policy.th --output mdp-complex/results.npz --meta-batch-size 1 --num-batches 2 --num-workers 2

Grad_Steps = 50 in test-my-new.py is the step you want to print. You can set it to any number. But you should keep Grad_Steps < num-steps (num-steps is set in mdp-complex/config.json)

We already save a trained model in mdp-complex.

1. maml_rl/envs/init.py register your environment:

   register( 'TabularMDP-v1', entry_point='maml_rl.envs.mdp-my:TabularMDPEnv', kwargs={'num_states': 5, 'num_actions': 3}, max_episode_steps=10 )

2. basic settings: /configs/maml/mdp/mdp-my.yaml

3. change the environment: /maml_rl/envs/mdp-my.py

4. train & test

Our basic algorithm is based on MAML:

https://github.com/tristandeleu/pytorch-maml-rl

MAML project is, for the most part, a reproduction of the original implementation cbfinn/maml_rl in Pytorch. These experiments are based on the paper

Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. International Conference on Machine Learning (ICML), 2017 [ArXiv]

Reinforcement Learning with Model-Agnostic Meta-Learning (MAML) Implementation of Model-Agnostic Meta-Learning (MAML) applied on Reinforcement Learning problems in Pytorch. This repository includes environments introduced in (Duan et al., 2016, Finn et al., 2017): multi-armed bandits, tabular MDPs, continuous control with MuJoCo, and 2D navigation task.

If you want to cite this implementation of MetaRLSAS:

@inproceedings{mingyue21ameta,
  author    = {Mingyue Zhang and Jialong Li and Haiyan Zhao and Kenji Tei and Shinichi Honiden and Zhi Jin},
  title     = {A Meta Reinforcement Learning-based Approach for Online Adaptation},
  booktitle = {{IEEE} International Conference on Autonomic Computing and Self-Organizing
               Systems, {ACSOS} 2021},
  publisher = {{IEEE}},
  year      = {2021}
}