A challenge in machine learning is to devise methods that allow
incorporating human insight into the automated learning process.
Current learning methods employ representations that make it
difficult to encode simplification and specific examples, and
learning is based on random exploration that is difficult to direct.
NEAT is a learning system where the learned decision policy is
represented in neural networks and learned through evolutionary
optimization, i.e. genetic algorithms. NEAT evolves network structure
as well as weights, which makes it possible in principle to
incorporate human guidance in three ways: (1) building a gradually
more complex network structure through shaping from simple to more
complex tasks, (2) training networks with examples of human behavior,
and (3) converting human-designed rules into network structures.
These techniques will be developed and evaluated in the domain of
designing complex behaviors for autonomous agents in the OpenNERO 3D
simulation environment. In a series of human subject experiments, the
solutions designed through human-guided neuroevolution will be
compared to those designed by human engineers and to those discovered
by neuroevolution alone, verifying that (a) the human-guided approach
results in better solutions, and (b) those solutions are more creative.
The result of this project is a machine learning approach will allow
engineers to generate creative designs to many real-world sequential
decision problems. Applications of this approach will lead to safer
and more efficient vehicle, traffic and robotic control, improved
process and manufacturing optimization, and more efficient computer
and communication systems. It will also make the next generation of
video games possible, with characters that exhibit realistic and
adaptive behaviors; such technology should lead to more effective
educational and training games in the future.
This research is supported by the National Science Foundation under grant IIS-0757479.