Swarm Control

Advances in control engineering have made it possible to manipulate the motion of man-made objects such as automobiles, hard disks, and drones with a high degree of precision and reliability. In the case of hard disks, for example, the readout device is controlled with nanometer-order precision. The concept of feedback control plays an indispensable role in the design of antilock braking systems, which are now indispensable in automobiles. Significant advances in the control of individual artifacts, both theoretical and applied, occurred in the 20th century.

On the other hand, the control of the movement of large numbers of nonartificial objects, especially biological populations, remains a difficult problem. Variations in individual characteristics, lack of reliable dynamic models, and non-negligible non-linearities make it difficult to approach the problem using conventional control engineering alone. Therefore, the Wakamiya Laboratory is developing innovative information and communication technologies for guiding biological populations by learning from living organisms while taking advantage of the strengths of control engineering.

Our research targets a wide range of biological communities, including livestock, fish, birds, and swarms. We develop information and communication technologies that are appropriate to the characteristics of each flock, the characteristics of its environment, the range of possible control strategies, and the objectives of guidance. We use both theory and simulation in a flexible manner and continuously improve our techniques by applying them to real communities.

Deep unfolding-based control

Although the optimal control problem of a nonlinear system is one of the most general forms of continuous optimization problems, it is generally impossible to obtain an analytical solution, and various numerical methods have been proposed. However, it is difficult for any of the existing methods to cover all forms of optimal control problems for nonlinear systems, and the need for advanced mathematical knowledge has been a major hurdle.

In this study, we propose a control system design method using deep unfolding, in which state changes of a dynamic system are expanded into a deep neural network with a multistage layered structure. In this neural network, each layer represents the dynamic system at a particular moment in time. Each layer also contains trainable parameters that determine the control inputs. Once the computational graph of the deep neural network is determined, standard deep learning techniques can be used to learn the parameters that determine the control inputs of this dynamic system. We are in the process of verifying the effectiveness and usefulness of this control method.