Skip to main content

Measuring prejudice in autonomous agents

Scroll down

Case study

Measuring prejudice in autonomous agents

Roger Whitaker, Professor of Collective Intelligence at Cardiff University and Academic Director of Supercomputing Wales, has been working in collaboration with MIT on research which suggests that prejudice can easily develop in populations of autonomous agents. This has implications for understanding prejudice, and the development of

prejudicial groups, in humans.

“We were trying to understand what makes simple agents or bots become prejudicial, and whether it is a phenomenon that exists in nature – are there pressures on a population to become prejudicial. So it’s quite abstract, and fundamental, but it requires a lot of supercomputing power,” Whitaker says.

Whitaker’s team set up a model to test a cooperation problem, where agents have to interact and decide if they’re going to help each other.

“It’s called indirect reciprocity and it’s seen a lot in everyday life. Holding a door open for someone, for example: there’s a small cost to us, but a greater benefit for the other person,” he says.

“We set this up as an artificial life problem where agents have to decide whether to donate to other agents. We set it up to examine whether agents become prejudicial: whether they would start to decide not to cooperate based on the recipient not being in the same group as the donor,” he says.

Using Supercomputing Wales, the team was able to run “thousands of agents playing together in simulated donation games where the agents had the opportunity to form groups based on prejudice”, he says.

“We were able to control the experimental parameters, and observe conditions where agents with similar prejudicial views formed groups, through what is called homophily. Agents with a similar prejudicial disposition were assumed to have a higher chance of assimilating with one another based on being comfortable with each other’s behaviour.

“We found that there are forces that seem to make it easy for individuals to become prejudicial and to come together in groups. Within those groups, cooperation emerges – so they become highly cooperating but isolated units,”

Whitaker says.

“It also showed that when you get systems of devices that have some degree of autonomy and sensing, there is the potential for prejudice to take hold of its own accord,” he says.

Working with MIT helped to raise the profile of the project, Whitaker says, and was a good chance to show what Supercomputing Wales offers.

“This international collaboration showed that Wales has cutting edge facilities and capabilities, including the capacity to accelerate research through highly skilled research software engineers. This environment allowed the work to reach its full potential” he says.

“These models are heavily dependent on random choices by agents. They’re also evolutionary: over generations, agents need to start interacting and then periodically stop and decide whether to change behaviour based on social learning. The supercomputer allows us to play out scenarios and run the same simulations again and again, with different random starting points. Our RSEs are able to help optimise that, and have a background to contribute to the simulations, understanding the concepts and suggesting how to translate them for maximum computational efficiency.”

Get in touch

Supercomputing Wales
Data Innovation Research Institute
Cardiff University
Trevithick Building
CF24 3AA