Skip to content

Latest commit

 

History

History
42 lines (27 loc) · 2.62 KB

README.md

File metadata and controls

42 lines (27 loc) · 2.62 KB

Function Approximation

Learning Goals

  • Understand the motivation for Function Approximation over Table Lookup
  • Understand how to incorporate function approximation into existing algorithms
  • Understand convergence properties of function approximators and RL algorithms
  • Understand batching using experience replay

Summary

  • Building a big table, one value for each state or state-action pair, is memory- and data-inefficient. Function Approximation can generalize to unseen states by using a featurized state representation.
  • Treat RL as supervised learning problem with the MC- or TD-target as the label and the current state/action as the input. Often the target also depends on the function estimator but we simply ignore its gradient. That's why these methods are called semi-gradient methods.
  • Challenge: We have non-stationary (policy changes, bootstrapping) and non-iid (correlated in time) data.
  • Many methods assume that our action space is discrete because they rely on calculating the argmax over all actions. Large and continuous action spaces are ongoing research.
  • For Control very few convergence guarantees exist. For non-linear approximators there are basically no guarantees at all. But they tend to work in practice.
  • Experience Replay: Store experience as dataset, randomize it, and repeatedly apply minibatch SGD.
  • Tricks to stabilize non-linear function approximators: Fixed Targets. The target is calculated based on frozen parameter values from a previous time step.
  • For the non-episodic (continuing) case function approximation is more complex and we need to give up discounting and use an "average reward" formulation.

Lectures & Readings

Required:

Optional:

Exercises