This section is about making decisions based on an automated processing of the available information.
Examples of such decisions are classification (medical diagnosis or fault detection and isolation) and automatic control (computation of input signals to impose some required behaviour on a system). Such decisions usually rely on mathematical models that may have to be built.
Principles and physical laws lead to white-box or knowledge-based models (e.g., in fluid mechanics, electrical engineering, chemistry, biology, etc.). Observed input-output behaviours alone lead to black-box models (the input-output data may correspond to actual experimentation or to simulations of complex, realistic knowledge-based models; one then speaks of computer experiments). All the possible grey-box models in-between can be used.
Active topics researches on modelling include model reduction, data reconciliation, model choice for a specific task, model-based control, etc. Thus modelling and parameter estimation, often called identification, are important ingredients. Also important is the ability to infer, from the available data and using the model, the values taken by variables that are not directly accessible to measurement but play a critical role in the decision process. Depending on the community, this is known as state estimation, or observation or software sensing. Because of the imperfect nature of a model and of unavoidable perturbations, control usually involves some feedback. Output information, collected via sensors and suitably processed, is used for that matter, to modify the input signals, as implemented by actuators, in order to improve the system performances. Ensuring the stability of a controlled system is definitely a critical issue. This decision stage may involve sophisticated mathematical strategies accounting for nonlinearities, constraints, timevarying parameters, delay effect, etc.
Optimal control is a natural approach when accurate models are available, and the objectives to be pursued easily translate into an optimality criterion (often the case in spatial control problems). In many cases, however, the situation is far less clear-cut. Robustness of performance in presence of the so-called reality gap (i.e. discrepancy between the actual system and its mathematical model), and adaptability of the control system to changing conditions, may then become more important than optimality.
The design of decision and control systems needs to rely on the most advanced approaches to artificial intelligence and machine learning, control theory and dynamical systems, optimisation, signal processing, optimum design, probability and statistics, best applied in cooperation. Digiteo is interested in all aspects of decision and control systems, with a strong emphasis on black-box modelling, parameter and state estimation, nonlinear dynamical systems, the robust control of uncertain systems, constrained control and trajectory generation, and control taking advantage of the laws of physics. Much is expected from the collaboration of Digiteo groups that attack the same problems with different tools and from different angles.
Here is a non-exhaustive list of examples :
- In nonlinear state estimation, specialists of control, signal processing and statistics need to work together to deal with nonlinear, uncertain dynamical models and produce estimators that take uncertainty explicitly into account, as well as the intended use of the estimate.
- Black box modelling, artificial intelligence, control theory and statistics each contribute their set of tools. Attempts to bridge gaps are encouraged within Digiteo.
- In control, it is often assumed that the model and state of the system are available at the outset. The interactions between modelling, state estimation and control in a context of a very partial knowledge on the system and of unavoidable perturbations need to be investigated. Biotechnology, for instance, calls for advanced control strategies coupled to identification techniques. In this case, many state variables are not directly measured, requiring the use of state estimators (soft sensors) to provide information to the control part.
- In the automotive field, improving vehicle control, reducing the emission of pollutants, or saving energy for electrical or hybrid vehicles requires robust control strategies that take advantage of models based on physics.
- In the energy field, stabilisation and optimisation algorithms are required to help improving resource sharing and network performance. Advanced control coupled to optimisation should improve energy saving in buildings, and the efficiency of fluid and thermal transport systems.
- Control devices may be embedded in the system to be controlled, distributed on a network, which may itself be used to control distributed objects. Interactions between control theory, telecommunications and computer science should therefore play an important role.
- Close interactions between physicists and engineers are often a good way to determine insightful control strategies taking advantage on the natural properties of the dynamics under consideration. Prime examples can be found in the sensors themselves, which form an essential part of sensing systems and hybrid systems, another major scientific theme in Digiteo (cf. Hybrid Systems and Sensing Systems).
Some teams that may not consider themselves as working on decision and control systems do deal with actual uncertain dynamical systems that may need to be modelled and on which decisions have to be taken. Other teams that develop control and decision methodologies do not have the equipment, skills and manpower required to deal with actual applications, e.g., in biology, communication networks, grid computing, physics, robotics. Digiteo will encourage both types of teams to benefit from common application projects.