强化学习之路(0)之缘由
开始接触强化学习是在回到所里开始找课题研究方向的时候,偶然查到伯克利的机器人视频,给我比较惊艳的感觉。陆续开始查找强化学习的一些资料。最近看到一篇有关强化学习的博文,瞬时,想把自己接触到强化学习的过程也做一个简单梳理,为接下来准备在强化学习方向完成毕业论文做一个初略安排。
有关强化学习的资料:
-
Tom M. Mitchell的《机器学习》最后一章讲reinforcement learning,算是入门
-
吴恩达的博士生Pieter Abbeel(现任教于伯克利,上面的机器人即是他们实验室),他的博士论文Apprenticeship Learning and Reinforcement Learning with Application to Robotic Control值得一看,还有其在伯克利的DRL课程
-
Richard S.Sutton and Andrew G.Barto的RL的经典之作Reinforcement Learning: An Introduction
-
Marco Wiering Martijn van Otterlo (Eds.)写的 Reinforcement Learning State-Of-the-Art
-
2015 nature上的 一篇关于 reinforcement learning 的综述: Reinforcement learning improves behaviour from evaluative feedback(Michael L. Littman)(链接还包括2015年机器学习专栏其他5篇综述,如下:
- Machine intelligence (Tanguy Chouard & Liesbeth Venema)
- Deep learning(Yann LeCun, Yoshua Bengio & Geoffrey Hinton)
- Reinforcement learning improves behaviour from evaluative feedback
(Michael L. Littman) - Probabilistic machine learning and artificial intelligence(Zoubin Ghahramani)
- Science, technology and the future of small autonomous drone(Dario Floreano & Robert J. Wood)
- Design, fabrication and control of soft robots(Daniela Rus & Michael T. Tolley)
- From evolutionary computation to the evolution of things(Agoston E. Eiben & Jim Smith)
应该关注的实验室:
- Knowledge-Based Reinforcement Learning
- MIT飞行控制实验室Scaling Reinforcement Learning Methods by Incremental Feature Dependency Discovery
- Reinforcement Learning and Control
代码工具框架:
-
最近找到一个PyBrain的python开源库,不错,准备前期就用它了
-
看Pieter Abbeel视频中提到的,他们的库Computation Graph Toolkit
-
RL-Glue framework
-
网友提到的:
一些相关会议:
(from reinforcement learning State of the art)
A large portion of papers appears every year (or two year) at the established top conferences
- artificial intelligence, such as IJCAI, ECAI and AAAI
- top conferences with a particular focus on statistical machine learning , such as UAI, ICML, ECML and NIPS.
- In addition, conferences on artificial life (Alife), adaptive behavior
(SAB), robotics (ICRA, IROS, RSS) and neural networks and evolutionary computation(e.g. IJCNN and ICANN) feature much reinforcement learning work.
workshop:
- The European workshop on reinforcement learning (EWRL)
- IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning (ADPRL)
competitions: