Millions of computer-generated entities that live and die by natural selection could reveal how our own culture and language evolve.
The software agents are part of a project called NEW TIES (New and Emergent World Models Through Individual, Evolutionary, and Social Learning), which draws on the expertise of five European research institutions to push computer simulation of artificial worlds further than ever before.
The joint computer project not only reproduces individual and evolutionary learning, but also social learning.
“Social learning is these guys telling each other what they learn on their own. One is learning about hot and cold and another is learning about soft and hard.
“They exchange knowledge and save effort,” explained project coordinator Gusz Eiben, a professor of artificial intelligence at the Vrije University, Amsterdam.
Understanding gleaned from such a project could advance machine learning for a range of applications. The learning software could guide exploratory or search and rescue robots that must cooperate to accomplish tasks in unknown environments.
The simulation computer project could even allow policy makers to test out new laws before carrying them out in real life.
The team of computer scientists, sociologists and linguists are creating a population of millions of unique entities that have the ability to pass on life-prolonging tips to their community. In the process, they may evolve their own language.
Computers Create Unique Beings
Each agent is randomly generated by a computer to possess a gender and different variations of life expectancy, fertility, size, and metabolism. The randomness of their programs allows each one to behave differently even when faced with the same set of circumstances.
The outcome of their actions — moving around, talking with another entity, and giving birth — burns fuel that can be replenished by finding the right food source.
Those who lack the wherewithal to survive risk certain death and the inability to propagate their genes.
A simple vocabulary of five to tens words, such as “food,” “near” and “agent” gives the entities the basics for communication.
Meanwhile, an algorithm enables two entities to agree on the meaning of new words and could allow the artificial beings to evolve a language.
The idea is to expose the agents to challenges and see how they adapt and develop their own world models.
Recently, Eiben and his team began running their first simulations using 1,000 to 3,000 agents to ask the question: Does individual learning compensate for bad genes? Eventually they plan to scale up to millions of agents.
The big challenge the team faces, said senior researcher Michele Sebag, an expert in artificial intelligence at the University of Paris-Sud, is tracking the behavior of each of the millions of agents.
“You can’t look at every agent individually. You have to have new facilities in data mining to understand what is going on in your population,” she said.
Following the rationale that “birds of feather stick together,” Gusz will be pinpointing and profiling agents who cluster together as well as tracking the locations of each agent as it moves over time.