Our insights into AAAI

We reveal the highlights and PROWLER.io's activities at this year's conference.

back to our blogs

Our insights into AAAI

We reveal the highlights and PROWLER.io's activities at this year's conference.

I am attending the upcoming AAAI conference which showcases the latest research in the field of artificial intelligence. Its broad scope means attendees are exposed to new developments and ideas not only in their own area of expertise but across a wide spectrum of related fields.

The paper presentations are a core component of the event and are grouped into thematic sessions. I am particularly excited to hear more about papers falling under the game theory and economic paradigms, the multi-agent systems, and the reinforcement learning conference sessions.

PROWLER.io’s research interests are aligned with the major themes of the AAAI conference. Indeed, probabilistic modelling, reinforcement learning and multi-agent systems form the foundation of our research, and our teams are collaborating to bring these fields together.

This inherent need to start connecting the many areas of artificial intelligence is also recognised at AAAI. One of its new workshops, Reinforcement Learning in Games, is dedicated to research at the intersection of reinforcement learning and game theory. This workshop is one of the reasons PROWLER.io could not miss the opportunity to take part at AAAI.

In the Reinforcement Learning in Games workshop, I will present our recent paper ‘Coordinating the Crowd: Inducing Desirable Equilibria in Non-Cooperative Systems,’  where we combine multi-agent reinforcement learning with black box optimisation. This work was led by David Mguni and Joel Jennings and a joint effort of PROWLER.io's multi-agent systems research team and Emilio Sison from MIT.  The same paper has just been accepted at the upcoming AAMAS 2019 conference and we also presented an application of the paper at the NeurIPS 2018 machine learning conference.

The paper tackles the problem of coordinating multiple selfish agents that act in a shared environment and, inevitably, affect each other's performance.

Let’s look at a real world example to understand the problem. In a city like London, there are a group of drivers on the road network that want to reach the same destination. For example, they all want to get to a business district like Canary Wharf. Each driver wants to minimise their journey time.

However, with every driver pursuing their own objective to get there as quickly as possible, the road network gets congested. Consequently, the group of drivers increase their journey time, spend longer on the road and air pollution also increases. This is because the selfish and uncoordinated actions of individuals adversely affect the journey times of other drivers.

To avoid this scenario, we propose a method in which an external party can incentivise agents to behave in a way that optimises system performance while taking into account their selfish behaviour. For example, during peak traffic times, the incentivising external party (which could be a traffic manager) could add a toll charge to certain roads, encouraging drivers to take different routes and reducing congestion. These tolls are calculated using a combination of black box optimisation methods and multi-agent reinforcement learning. Crucially, these incentives must be calibrated to avoid any undesired side effects. We don’t want to cause congestion in another area of the city.

Other than at the RLG workshop, I’m also presenting the paper ‘Partial Verification as a Substitute for Money’ at the main conference, which is a collaborative piece of work with Ian Kash from the University of Illinois at Chicago and Rafael Frongillo from the University of Colorado Boulder. In this paper, we propose a general framework that prescribes how to adapt existing mechanisms that incentivise agents to behave in a desired manner where monetary payments are accepted, to scenarios in which payments are not allowed.

AAAI is a great opportunity to interact with other researchers, get feedback on our work and attend a range of exciting presentations. Moreover, at this year’s conference, a town hall meeting is also planned to discuss a 20-year roadmap for AI research, which should generate a lively debate.

I am looking forward to seeing what lessons and insights this year’s conference brings to the dynamic world of artificial intelligence.

Join us to make AI that will change the world

join our team