Flanders AI Research Program opens new ways to solve societal and industrial challenges with Multi-Agent AI Systems
Industrial and societal problems are complex. Just think of the fight against epidemics or the transition to green energy. Researchers from VUB, UHasselt and UAntwerpen are developing new solutions using multi-agent systems as part of the Flanders AI Research Program. What are multi-agent systems? How do they work? In which applications do they already achieve results today?
What is multi-agent AI?
Today, neither people nor computer systems standalone. They are connected. They communicate and interact. Multi-agent AI systems consist of multiple autonomous software or robotic components (also known as agents) that work together to perform complex tasks.
For example, intelligent traffic lights that work together to make the flow of traffic in a neighborhood more dynamic and efficient. On a larger scale, multi-agent systems synchronize energy consumption and production in a region. In the future, agents that control wind turbines at sea, for example, will have to coordinate with other agents in the power grid. Their decisions and actions will ultimately influence the smart home agent that manages the charging of your electric car.
Agent & Agentic AIBy the word 'agent' we mean an entity that can act and thus influence its environment. This could be a human being, but in the context of AI it tends to refer to software and/or robotic components that act autonomously. An example of an agent is an AI system that manages a wind farm, continuously determining which turbines to turn off, which to run, and so on, based on measurements. Popular chatbots like ChatGPT don't fall under the term 'agent' because they are completely reactive: the user asks a question and the bot answers. No further interaction unless the user asks a follow-up question. That's why there's a lot of talk today about 'agentic AI' as the next step in the AI revolution: AI systems that connect to the 'real world' by autonomously solving complex problems that require multiple actions. So, they are 'agents'. |
Desired features of multi-agent systems.
Coordination between agents in a multi-agent system is essential. Coordination through a central system is however undesirable: it is not scalable, it compromises privacy by concentrating all data in one place, and, usually, no single agent has an overview of the whole system or has control over other agents. For example, the agents involved in making decisions about the energy production of a wind turbine are different from those supervising charging your car. Coordination between the agents in a multi-agent system must therefore be (implicitly) organized through communication, agreements or incentive mechanisms. Furthermore, agents in a multi-agent system should be able to adapt both to their changing environment and to sometimes hard to predict human behavior. They must therefore be able to learn from their experiences.
Three hundred professors, doctoral students and other researchers from eleven knowledge institutions work within the Flanders AI Research Program. Funded by the Flemish government, they are joining forces to tackle generic AI research challenges and their application in healthcare, energy sector, industry and society. The research on multi-agent AI systems is tackled within the Situated AI research challenge of the program. |
Researchers use concepts from game theory and simulations in multi-agent systems
AI engineer Marjon Blondeel is working on the Self-Sustainable Smart Grid use case at the Vrije Universiteit Brussel. She models the smart energy network of the future. "Traditional machine learning, classical software development and humans cannot cope with the complex challenges of regulating highly fluctuating energy production and consumption in the context of the current electricity grid," says Blondeel. "The solution is an intelligent power grid that learns when to redistribute power to match supply and demand. In such a grid, smart agents continuously make decisions, consult with each other and adapt to each other to create the best possible energy grid."
The research in the Smart Grid use case is based on a common theoretical framework that brings together several disciplines. "One of the PhD students working on this use case is using concepts from game theory. This allows us to zoom in on the strategic interactions between different agents and study the convergence of multi-agent systems. In addition, this method allows analysis in situations where large samples are not available," explains Ann Nowé. She is professor and head of the Artificial Intelligence Lab at VUB, one of the research groups active in the Flanders AI Research Program.
"Abstract game situations also teach us a lot about the interactions in our society," she says. "For example, they teach us why it is so difficult to get different parties to work together to combat climate change." Each party’s efforts must be sustained long enough to reap the benefits. And in the short term, it may be more advantageous for an individual agent not to make the effort, but to surf along with the efforts of others, which is called free-riding. Within game theory, several prototypical game situations have been described that model real situations. On the one hand, we can use mathematical analysis and simulations to study how such a situation will develop if each of the agents act in their own interest. On the other hand, we can design the rules of the game to make the desired collective behavior emerge from the individual agents’ best interest. For example, by introducing road tax pricing, where the question is what road taxes we should impose to achieve maximum traffic flow on a road network.
How does research evolve into industrial and societal solutions?
Within the Flanders AI Research Program, we conduct demand-driven research aiming for new industrial applications. Several use cases already show the potential of autonomous agents with flexible coordination mechanisms.
VUB professor Jan Helsen explains the use case for monitoring wind farms. "We have started collaborations on this topic with Parkwind and Norther, among others. The 800 sensors on the latest offshore wind turbines transmit data every second. This data is analyzed by artificial intelligence. The analysis tells us how the wind turbines are performing, under what conditions and in what weather conditions they are starting and stopping, and how the turbines at the back of the farm are performing differently than the turbines at the front, which have a free supply of wind".
The results of AI to monitor and control wind farms are impressive. Data and multi-agent AI systems appear to be the key to higher yields and lower maintenance costs for these green power plants. "If the load on the wind turbines is distributed based on an analysis of the data using AI, a damaged turbine can be relieved by turning it off or letting it turn at lower speeds.[SD3] Optimized control of the park allows operators to group all the necessary maintenance work together and carry it out in one go in the summer, when the weather is favorable. The cost of AI-driven maintenance is only a fraction of what traditional proactive maintenance and emergency repairs to faulty turbines cost and eliminates unnecessary inspections and replacements.”
"When building a multi-agent system, it is important to realize that each agent takes actions according to its own internal logic and to then anticipate the resulting dynamics at the system level," says Ann Nowé. "It doesn't make sense to make one traffic light intelligent, if traffic is stuck at the next crossing. No, you have to make a whole intersection smart, a whole street, a whole neighborhood. Moreover, the system has to react intelligently to changing traffic flows because better flows might attract more traffic.
The same applies to dynamic pricing of green electricity. You can use the price to achieve distributed coordination between agents, since consumption patterns will change when you allow the energy price to vary based on supply and demand. Thus, when little electricity is being produced and much is demanded, the price will climb and discourage consumption. However, such dynamic pricing needs to be done carefully, as low energy prices might still lead to overconsumption: for example, if all households start their washing machines or charge their cars when energy prices are low, consumption will spike and energy production might still not meet demand. So, either the price setting agent needs to be able to anticipate the reaction of the households, or other incentive mechanism must be used. A further challenge is that energy price is not the only metric consumers might consider to make decisions. Comfort is another one, as for example many people will require their electric cars to be charged by a certain time of the day, even at higher prices. This means that incentive mechanisms must be carefully thought through, to be able to deal with dynamic interactions, including human behavior.”
We also use agent-based systems to address health-related challenges. "With epidemiological models, we can simulate epidemics and interventions. By introducing agent-based AI techniques, both for simulating epidemics and for learning from interventions, for example through reinforcement learning, we can learn, in close collaboration with epidemiologists, how to better tackle future epidemics," says Prof. Pieter Libin of the VUB AI Lab. "This is interdisciplinary work that brings together the expertise of researchers from VUB, UHasselt and UAntwerp." He points out that learning such strategies is very complex: "After all, there are a lot of conflicting goals that need to be taken into account, such as the pressure on hospitals, but also the impact on society by introducing restrictive measures. We therefore research AI-based decision-support systems that assist experts in their complex decision-making tasks. These systems must balance different objectives and aim for a fair distribution of the effects of the measures taken.” Prof Libin concludes: "Since we want to support human decision-makers, it is very important to learn their various perspectives from feedback and take them into account when presenting them with several distinct trade-off solutions".


