16 Feb 2020 -

Universal Value Function Approximator

This blogpost explores an implementation of UVFA and evaluates it on a few gridworld tasks.

Gridworlds with a single goal represent easy challenges and can be solved with tabular methods. But what if we don’t want to use a fixed goal? UVFA presents a method to learn a general representation of value functions that should generalise to new goals not present in the training levels.