From Data to Interventions: Using System Identification and Robust Control Algorithms to Design Effective Treatments

Open Access
Bekiroglu, Korkut
Graduate Program:
Electrical Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
June 17, 2015
Committee Members:
  • Constantino Manuel Lagoa, Dissertation Advisor
  • Constantino Manuel Lagoa, Committee Chair
  • Minghui Zhu, Committee Member
  • Kenneth Jenkins, Committee Member
  • Stephanie Trea Lanza, Committee Member
  • Necdet S Aybat, Special Member
  • Adaptive Intervention
  • Robust Treatment Design
  • System Identification
  • Atomic Norm
  • min-max Structured Robust Optimization
Behavioral and social scientists have demonstrated the advantages of the adaptive treatments, which usually provide better results than the fixed treatment (all patients get same type and level of treatment). Therefore, in this dissertation, we initially illustrate how one can develop dynamical models with suitable uncertainties for behaviors and identify them. Then we also demonstrate the usage of control engineering methods, such as feedback or adaptation, and robust optimization, to develop a systematic way for designing robust personalized treatment algorithms. This new robust adaptive treatment design consists of three steps. In this dissertation, we develop different algorithms for first and second steps. For the first step, three different identification algorithms (identification with Lasso, Parsimonious model identification of single input single output systems, and Parsimonious identification of multi input multi output systems) are proposed which can utilize intensive longitudinal behavioral data to identify model parameters, interpolate the missing data, and quantify the uncertainties in the model. Then, for the second step, we provide a detailed step-by-step explanation of how control engineering methods can be used to design a robust adaptive intensive intervention. Finally, the methods are evaluated via simulation. The performance of identification algorithms is demonstrated with synthetic behavioral data. The simulation results illustrate how the designed robust adaptive intensive intervention can produce improved outcomes with less treatment by providing treatment only when it is needed. The methods are robust to model uncertainties as well as to the influence of unobserved causes. As a result, these new methods can be used to design robust adaptive interventions that function effectively yet reduce participant burden.