| Summary: | Can symmetry be utilized as a design principle to constrain
evolutionary search, making it more effective? This dissertation aims
to show that this is indeed the case, in two ways. First, an approach
called ENSO is developed to evolve modular neural network controllers
for simulated multilegged robots. Inspired by how symmetric organisms
have evolved in nature, ENSO utilizes group theory to break symmetry
systematically, constraining evolution to explore promising regions of
the search space. As a result, it evolves effective controllers even
when the appropriate symmetry constraints are difficult to design by
hand. The controllers perform equally well when transferred from
simulation to a physical robot. Second, the same principle is used to
evolve minimal-size sorting networks. In this different domain, a
different instantiation of the same principle is effective: building
the desired symmetry step-by-step. This approach is more scalable
than previous methods and finds smaller networks, thereby
demonstrating that the principle is general. Thus, evolutionary
search that utilizes symmetry constraints is shown to be effective in
a range of challenging applications. === text
|
|---|
