Integrated learning and reasoning in intelligent systems
Mehdi Dastani is full Professor and chair of the Intelligent Systems group at Utrecht University. ‘What fascinates me most is the way intelligence emerges from the interplay of high-level cognition and low-level sensory perception.’
How did you end up in artificial intelligence research?
‘When I came to the Netherlands back in 1985, computers were beginning to enter our daily lives. Since I was fascinated by computers and the possibilities they offered to automate tasks, I enrolled in a computer science program at the University of Amsterdam. By that time, artificial intelligence was an upcoming field that was taught as a special track with electives such as robotics, logics, expert systems, and image analysis. I got intrigued by the fundamental questions people were asking about the nature and feasibility of such a thing as “artificial” intelligence. Since I wanted to dive deeper into these fundamental questions, I decided to pursue a second master’s in philosophy.’
What types of research have you been conducting over the years?
‘In the early nineties there weren’t that many data, the available computing power was far from sufficient, and the data-driven techniques were not as advanced as today. At the time, the artificial intelligence community was mainly working on model-driven approaches using symbolic knowledge combined with deliberative reasoning processes. Over the years, along the developments in the field, digitalization and datafication of society, my research slowly migrated towards a combination of model- and data-driven approaches. My current research focuses on the design and development of autonomous systems of which their behaviors are based on reasoning and learning.’
What is the biggest scientific challenge in your field?
‘There are many great challenges in Artificial Intelligence, but what fascinates me most is the way intelligence emerges from the interplay of high-level cognition and low-level sensory perception. For example, when your eyes see a coffee cup, you jump immediately to the conclusion that it should be handled with care, since it’s probably fragile. How to model and couple cognition and perception in artificial intelligent systems in a systematic and principled manner is currently a major challenge in artificial intelligence.’
What kinds of projects are you currently working on?
‘In my research group, we are exploring several research themes in the area of knowledge representation and reasoning, multiagent reinforcement learning, agent-based modelling, and causal inferencing. We also collaborate with external partners on real-world projects related to these research themes. For example, we did a project in The Hague Southwest, which is a district with some 80,000 inhabitants. The municipality wanted to introduce call-buses to reduce car use, and was eager to understand how residents would respond to this new mobility service before implementing it and making investments. To study the possible effects of this mobility service, we modelled the real population using existing datasets and our home grown agent-based modelling techniques, creating 80,000 synthetic individual agents that reflect the real demographic properties and mobility behaviors of the residents. The synthetic population was then used to simulate the response of the residences to the new mobility service and to identify key factors that would incentivize its adoption. This project followed a previous one in which we investigated the spread of COVID-19 in the US state of Virginia with 3 million residents.
What is you ambition with your research?
‘Application oriented projects like the ones I mentioned are nice because they contribute to societal problems, but such projects need to be based on the results and insights gained from basic research. How can we model intelligent systems that react to and collaborate within social settings? How do they collaborate when their observations are limited? And how can their behaviors be influenced through interventions in their social settings? My current research focuses on developing a generic framework that integrates machine learning techniques with deliberative reasoning in a systematic and principled manner. This integration is crucial for designing intelligent systems that can learn and reason while operating in dynamic social environments, where other intelligent systems, including humans, are also active.’
As of 2024, you have become the representative for IPN’s Special Interest Group Artificial Intelligence. Why did you decide to take up this role, and what do you hope to achieve?
‘One of the things I am committed to, is to strengthening the connections between the Dutch AI community and IPN. As a first step I have installed an advisory board composed of four AI researchers from different institutions. We plan to bring the community together during a dedicated track at ICT.OPEN, and preferably also at other annual events, for example BNAIC/BeNeLearn. Artificial intelligence is a broad interdisciplinary research field. While the focus of IPN is on computer science, AI involves other scientific disciplines such as social sciences and humanities. This can sometimes create communication challenges. As a Special Interest Group of IPN, we aim to bridge this gap by connecting the AI community with the computer science community at a national level.’
Foto: Sjoerd van der Hucht