Black-Box Optimizations in RL
In this chapter, we will change our perspective on reinforcement learning (RL) training again and switch to the so-called black-box optimizations. These methods are at least a decade old, but recently, several research studies were conducted that showed their applicability to large-scale RL problems and their competitiveness with the value iteration and policy gradient methods. Despite their age, this family of methods is still more efficient in some situations. In particular, this chapter will cover two examples of black-box optimization methods:
-
Evolution strategies
-
Genetic algorithms
