While Evolution Strategies (ES) are widely known as one of the streams of
evolutionary computation and are meanwhile regarded as a competitive
alternative to standard learning techniques in Reinforcement Learning,
most people associate ES with the covariance matrix evolution strategy.
However, there is more than just this particular evolutionary algorithm
designed for unconstrained real-valued search spaces.

This introductory tutorial provides a broader perspective and view stressing
the design philosophy of Evolution Strategies being not restricted to a
specific search space, such as unconstrained real-valued optimization, but
also includes discrete and combinatorial search spaces. This philosophy can
be best understood from the ES history that started from the evolution of
material objects - nowadays often referred to as hardware-in-the-loop
evolutionary optimization. That is, evolution is done on the "phenotype."
Accepting the constraints involved in such optimizations, one naturally can
derive design principles for mutation and recombination operators and the
control of those operators by self-adaptation - one of the great inventions
of ES. Special emphasis is put on a vivid understanding of how the ES
explores the search spaces. Recent findings will be presented and supported
by live experiments to explain the ES's ability to locate global optima in
landscapes with a huge number of local optima. The tutorial will also
investigate the reasons why ESs are now regarded as a scalable alternative
to Reinforcement Learning.

The tutorial will include a live computer experiment demonstrating the
relevance of the design and working principles discussed. This tutorial will
be on an introductory level requiring only a minimum of maths.