Co-evolution in predator prey through reinforcement learning
Authored by Megan M Olsen, Rachel Fraczkowski
Date Published: 2015
DOI: 10.1016/j.jocs.2015.04.011
Sponsors:
Clare Boothe Luce Program
Platforms:
Java
MASON
Model Documentation:
Other Narrative
Flow charts
Mathematical description
Model Code URLs:
Model code not found
Abstract
In general species such as mammals must learn from their environment to
survive. Biologists theorize that species evolved over time by ancestors
learning the best traits, which allowed them to propagate more than
their less effective counterparts. In many instances learning occurs in
a competitive environment, where a species evolves alongside its food
source and/or its predator. We propose an agent-based model of predators
and prey with co-evolution through linear value function Q-learning, to
allow predators and prey to learn during their lifetime and pass that
information to their offspring. Each agent learns the importance of
world features via rewards they receive after each action. We are
unaware of work that studies co-evolution of predator and prey through
simulation such that each entity learns to survive within its world, and
passes that information on to its progeny, without running multiple
training runs. We show that this learning results in a more successful
species for both predator and prey, and that variations on the reward
function do not have a significant impact when both species are
learning. However, in the case where only a single species is learning, the reward function may impact the results, although overall
improvements to the system are still found. We believe that our approach
will allow computational scientists to simulate these environments more
accurately. (C) 2015 Elsevier B.V. All rights reserved.
Tags
Simulation