Introduction to Reinforcement Le
2017-01-03 本文已影响66人
威仪棣棣
- Goal: Learn from reward to adapt the environment
- Setting:
- action/decision(agent -> environment)
- reward/state (environment ->agent)
- 怎么理解state？ -> 人对狗狗施加的命令
- Policy -- for Agent: learning a classifier(state->action)
- Agent's Goal: discounted reward $\sum_{t=1}^\infty \gamma^t r_t$
- Difference between RL and planning:
- RL: learning a model and find policy from samples
- Planing: find an optimal solution with a well-defined problem.
- Difference between RL and SL:
- All learn the model, but SL学的是batch的数据，一次性学，从数据到算法到模型，DAG单向路径
- RL：闭环/数据不同，环境->数据->算法->模型->环境
- 决策影响实践/环境 ： 强化学习适用
