Power cuts, financial network failures and supply chain disruptions are just some of the many of problems typically encountered in complex systems that are very difficult or even impossible to control using existing methods. Control systems based on artificial intelligence (AI) can help to optimise complex processes – and can also be used to develop new business models. Professor Lucas Böttcher, Professor of Computational Social Science at Frankfurt School, together with Dr Nino Antulov-Fantulin, Senior Scientist at ETH Zürich, and Thomas Asikis, doctoral student at ETH Zürich, have just published a paper presenting AI Pontryagin, a versatile AI-based control system designed to steer complex systems and networks towards desired target states. Using a combination of numerical and analytical methods, the researchers demonstrate how AI Pontryagin automatically learns to control systems in near-optimal ways even when the AI has not previously been informed of the ideal solution. The paper was published in the renowned scientific journal Nature Communications.
Fluctuations in complex systems are capable of triggering cascades and blackouts. To avoid such incidents and improve resilience, system specialists have devised a wide variety of control mechanisms and regulations; typical applications include voltage control in power grids, for example, or stress testing in financial institutions. And yet it is not always possible to control complex dynamic systems by manual intervention. In their paper, the researchers show how AI Pontryagin automatically learns quasi-optimal control signals for complex dynamic systems. The researchers’ analysis lays much of the vital groundwork; further research is still required to determine the system’s applicability to specific, real-world cases. At present, control methods are typically used to, for example, protect power grids from fluctuations and outages, manage epidemics, and optimise supply chains.
To use AI Pontryagin as intended, the AI must first be provided with information on the target system’s dynamics. In supply chains, this might include details of the number of possible suppliers, as well as purchasing costs and turnaround times. This information is used to determine which areas require dynamic optimisation. Users must also provide information on the system’s initial status, such as current stock levels, and its desired (target) status, such as the requirement to replenish stock to certain levels while minimising the use of resources.
The paper is available for download here.
For more information on AI Pontryagin, see Professor Böttcher's post on the Frankfurt School Blog here.