Abstract
In this paper we use the genetic programming technique
to evolve programs to control an autonomous agent
capable of learning how to survive in a hostile
environment. In order to facilitate this goal, agents
are run through random environment configurations.
Randomly generated programs, which control the
interaction of the agent with its environment, are
recombined to form better programs. Each generation of
the population of agents is placed into the Simulator
with the ultimate goal of producing an agent capable of
surviving any environment. The environment that an
agent is presented consists of other agents, mines, and
energy. The goal of this research is to construct a
program which when executed will allow an agent (or
agents) to correctly sense, and mark, the presence of
items (energy and mines) in any environment. The
Simulator determines the raw fitness of each agent by
interpreting the associated program. General programs
are evolved to solve this problem. Different
environmental setups are presented to show the
generality of the solution. These environments include
one agent in a fixed environment, one agent in a
fluctuating environment, and multiple agents in a
fluctuating environment cooperating together. The
genetic programming technique was extremely successful.
The average fitness per generation in all three
environments tested showed steady improvement. Programs
were successfully generated that enabled an agent to
handle any possible environment.
Users
Please
log in to take part in the discussion (add own reviews or comments).