Effects of diversity

PhD student: Chris Bennett

Supervisory team: Jonathan Lawry and Seth Bullock

Same difference?

Multi-agent autonomous systems usually involve agents with different characteristics. This includes human-machine teams or where agents have very different functions, but also where autonomy is introduced gradually, such as in automated vehicles on road networks. Even swarms of nominally identical agents will develop diversity as agents deteriorate or learn individually in response to different experiences of the world.

Understanding these effects would also allow the design of heterogeneity in multi-agent systems to be explored.

Diversity in multi-agent systems is often seen as a strength in biological systems but is generally avoided in engineered systems as it creates uncertainty in the detail of how the system will operate.

Our research investigated whether diversity could predict the performance of a multi-agent system or if additional information was required.

Research findings

We used computer simulations of a diverse swarm of agents being herded by two ‘shepherd’ agents. Diversity and behaviour were created by:

• Dividing the swarm into multiple social groups
• Setting whether agents preferred to be with their own or another social group
• Setting the strength of the group preferences

This simple system created diversity in the population and explored different strengths and types of interactions between agents. Performance was measured by how quickly and efficiently the herding task was completed.

The outcome of the simulations is clear: an increased number of social groups (increased diversity) degrades performance when agents prefer to be with their own group but improves performance when agents prefer to be with a group other than their own. This shows that it is not just diversity that controls performance in multi-agent systems, but also the nature of interactions between agents.

Implications: task, diversity, interactions

Diversity in multi-agent systems cannot be avoided and can sometimes even be beneficial. Our work shows that understanding the impacts of agent diversity in a particular context must also consider agent interactions. To do otherwise will result in an oversimplified view that can often be misleading.

This has implications for many scenarios:

• Road networks: The effect of different levels of vehicle autonomy on road networks is challenging to address but understanding must include how vehicles sense and respond to each other.
• Internet of Things: The diversity of connected devices that forms the Internet of Things means that understanding device interactions is essential to designing optimum data protocols.
• Swarm behaviour: Understanding how degradation of robot agents in swarms affects their system behaviour must acknowledge the gradual diversification in agent interactions.
• Ecosystems: Our approach could offer insight into how interactions between biological systems and species might impact an ecosystem’s ability to withstand shocks and adapt to external pressures.

Whatever the task, diversity in multi-agent systems remains important, but considering agent interactions is also critical to predicting the system behaviour.