Technology Research News, January 2, 2002
Software agents evolve purpose
By Kimberly Patch, Technology Research News
Behavior has always been a touchy subject, especially when academics delve into just how much of what we do is conscious choice and how much of it is instinct born of evolution.
Researchers from the Keldysh Institute of Applied Mathematics in Russia have shown that purposeful behavior, or motivation can emerge naturally in a software simulation that has simple software beings, or agents, evolving over many generations.
The researchers' simulation showed that a system that uses motivations to control simple reflexes can emerge in an evolutionary process, said Mikhail Burtsev, a researcher at the M. V. Keldysh Institute for Applied Mathematics at the Russian Academy of Science in Russia.
Having motivation was an advantage likely to be passed on to subsequent generations of the agents, he said. "The population of agents with motivations had obvious selective advantages compared with the population of agents without motivations," he said.
The research is ultimately pointed toward better understanding cognition in order to develop more realistic agents, Burtsev said.
The simulation provided the software agents with a simple world of 900 discrete pens where patches of grass grow. The agents lived in the pens and ate the grass.
With the energy they gained from the grass, the agents could do the following things: move to a neighboring cell on the left or right or jump over several cells in a random direction, eat, and mate. The agents could also choose to rest, saving energy.
When an agent ran out of energy, however, it died. Mating sometimes resulted in the birth of new agents.
The actions of each agent were determined by it's neural network, a set of logical neurons with connections to other neurons. The strength of these connections determined which action an agent took in a given situation. As a population of agents evolved, the strengths of the neural connections changed, modifying the behavior of the agents, said Burtsev.
The agents had several initial instincts: if an agent saw grass in its own cell it ate the grass, if an agent saw grass in a neighboring cell it moved into that cell, and if an agent saw another agent in one of the neighboring cells it tried to mate with its neighbor. If things got crowded and an agent saw two agents in both neighboring cells, it jumped to a new place.
The researchers populated the pen world with 200 identical agents. During the simulations, which took the agents through several thousand generations, their number varied from several dozen to as many as 800.
The agent population as a whole had one goal -- survival. This goal required individuals to push toward two basic subgoals -- to replenish energy, and to reproduce, said Burtsev. The agents evolved to seek out grass and other agents.
"The most important thing here is that we didn't force agents to follow these needs. The needs were prescribed explicitly by [the] environment, and only agents that had these two needs could successfully undergo selection pressure," said Burtsev.
The researchers concluded that the environment was the cause of the motivation -- at least in this case. "In the context of our model we can say that purposeful behavior is behavior directed to get those subgoals," he said. "These goals may seem trivial, but they're important features of all living creatures," he said.
Because the agents evolved this type of system, it must be more effective than behavior governed only by simple reflexes, Burtsev added.
The researchers also found that the agents evolved more intelligent behavior as time went on. Initially, the main action of the agents was eating. Later generations rested more, however. "This means that the agents became able to do only the right actions at the right time. We didn't foresee such intelligent strategy, and this was surprising," Burtsev said.
The work combines and improves several other artificial life approaches, and shows positive results regarding the usefulness of motivation, said Marcus Hutter, a researcher at the Research Institute for Artificial Intelligence in Switzerland.
However, the researchers' conclusions about motivation are tenuous, he said. "I think the motivated system performs better just because its neural net has access to additional useful information. It is useful to verify that this works, but the interpretation of the authors [that this is due to] additional motivation inputs is not convincing," he said.
The researchers plan next to create a more complex model that would force the agents to develop a hierarchy of goals, said Burtsev.
The model could be adapted to provide agents that could be used on the Web within five to seven years, Burtsev said.
Burtsev's research colleagues were Vladimir G Red'ko and Roman V. Gusarev of the Keldysh Institute of Applied Mathematics. The research was funded by the Russian government.
Timeline: 2-3 years, 5-7 years
TRN Categories: Artificial Life; Evolutionary Computing; Applied Computing
Story Type: News
Related Elements: Technical paper, "A Life Model of Evolutionary Emergence of Purposeful Adaptive Behavior," posted on CORR at http://xxx.lanl.gov/abs/cs.NE/0110021.