Laboratory for Information and Decisions for Complex and Uncertain Systems [Home]

1. DATA-BASED MODELING & CONTROL

System identification in the presence of non-stationary disturbances

Disturbances occurring in most chemical processes exhibit non-stationary behavior. On the other hand, almost all identification methods currently used assume a stationary disturbance behavior, either explicitly or implicitly. The consequence is that these methods can behave very badly when the non-stationary characteristics are significant. In this research, we are studying the implication of non-stationary disturbances for some popular identification methods and are developing some new methods that can handle this type of disturbances more effectively. Non-stationary characteristics can range from simple mean shifts (e.g., Wiener) to more extensive changes in the correlation behavior. Specifically, we explore the use of Markovian Jump Linear Systems (MJLS) as an attrative model structure for describing disturbance patterns such as intermittent drifts and abrupt jumps.

member(s) involved in this research area: Wee Chin Wong

Model free control of Non-Linear Systems

Traditionally, modeling, and optimal control have been researched as independent domains. The boundaries drawn among them are mostly for academic and historical reasons and are artificial to practitioners who are faced with the daunting task of integrating the various inconsistent tools and theories.
Clearly, modeling and control must play complementary roles. This is because one would seldom have the luxury of a complete, perfect model (giving accurate predictions on its own) or accurate measurements of all the important variables. produces based on them. The interplay between model / sensor accuracy and the closed-loop performance of a control algorithm can be subtle and intricate - enough that a model, a set of sensors, and a control algorithm that may be judged "good" in their own rights may compose a poorly-performing model-based control system.

As a solution, we propose a model-free solution via Q-learning, which can be viewed as a direct form of adaptive control. Specifically, we are working on algorithms on the model-free regulation of Markovian Jump Linear Systems with partial state-feedback.

member(s) involved in this research area: Wee Chin Wong

2. MULT-STAGE DECISION MAKING UNDER UNCERTAINTY

Managers and executives are always faced with the problem of making good decisions in the face of future uncertainty. These issues occur in scheduling and supply-chain problems, and portfolio management. We are interested in incorporating notions of risk in the methods we develop. Also, we are working on novel problems (such as supply chains faced with perishable resources) which fall under the same framework.

member(s) involved in this research area:
Nikolaos Pratikakis; Rakshita Agrawal; Farminder Anand

3. IDENTIFICATION OF COMPLEX NETWORKS

We are interested in identifying the governing relationships between the nodes of a complex network. The major difficulty lies in the fact that data is usually scarce. Interesting applications include, but are not limited to, gene regulatory networks.

member(s) involved in this research area: Ugur Guner