Plan

[chenkaiyu1997]

Four keypoints:

1. Learning Dynamics.
   • Firstly try: Ensemble Dynamics using MSE. (MB-MPO)
   • Then, try RNN.
   • Model uncertainty: MDN or GANonZ , on 1234 dataset.
   • Afterwards: Use Stacked Images.
   • Multi-level Long-term Dynamics.
2. Planning
   • MPC
   • MCTS
3. Curiosity & Exploration
   • Ensemble Value Functions (RPF)
   • RND
4. Hierachy (optional)
   • Tree-like Option Discovery & use and disuse
   • DAIYN

[chenkaiyu1997]

1. Add MPO, MPC, RPF & Test. DDL1.6 1.5 Start writing a paper.
2. Try RNN-MDN or GAN/NCE on 1234 Dataset, and Multi-level Long-term Dynamics.
3. If 2 ok, then add RNN-MDN/GAN/NCE & Test.
4. Try Tree-like Option Discovery & use and disuse, or DAIYN.

[zdx3578]

Ensemble Value Functions (RPF)：Randomized Prior Functions for Deep Reinforcement Learning ref by polo and openai RND; https://blog.openai.com/reptile/ meta learning; https://github.com/openai/supervised-reptile https://github.com/tristandeleu/pytorch-maml-rl 代码不少

https://github.com/eambutu/snail-pytorch https://github.com/sagelywizard/snail

https://github.com/thanard/me-trpo https://sites.google.com/view/mb-mpo/code

infobot(IB) code will modeb base: EMI(IB)

sac +HER . polo--rpf ; her：目标

infobot . goal ref : HER, Unsupervised Meta-Learning for Reinforcement Learning 意图-无监督-LEARNING A PRIOR OVER INTENT VIA META-INVERSE REINFORCEMENT LEARNING https://arxiv.org/abs/1805.12573v3