Some contributions to decision-making problems

Le résumé fourni par la source

This thesis, motivated by applications in the industrial and health sectors, is a collection of studies on different decision problems.In the first part, we focus on single-step decision-making problems where a predictive model is used upstream of the decision-making, and where explicit feedback is received. We propose to leave the task of defining the loss function associated with the predictive model to the end-user, in order to encode the real cost of using a forecast to make a decision. As an algorithmic approach, we consider decision trees, optimized with the adjusted loss function, and function approximation methods, similar to Q-value function approximation in reinforcement learning, in case where only the immediate reward is of interest. Three applications are studied: the calibration of an alarm system to fight against medical wandering; the problem of nomination in an electricity market, from the point of view of a renewable energy supplier; the optimization of production in the uncertainty of customer demand.In the second part, we focus on two specific problems of sequential decision-making, which we address using a Markov decision-making framework and reinforcement learning algorithms. In the first application, we try to optimize meal timing and insulin management for people with type I diabetes who rely on self-injections. To do so, we rely on a patient simulator, which is based on medical knowledge of the interaction between glucose and insulin and on physiological parameters specific to the patients. In the second application, we try to build an adaptive predictive questionnaire for smooth interactions with users. For binary data, the questionnaire looks like a decision tree, optimized in a bottom-up way. For non-binary data, this new questionnaire only asks questions that have already been asked, remembers previously observed values, and exploits them fully once they arrive in a terminal node, where a specific prediction function is available.In our final section, we look at three decision processes that, by construction, do not require the agent to explore the environment. For example, we consider a system whose dynamics are sufficiently stochastic that, whatever our action, we explore the state space, while having some influence through our actions. We also consider a system where some actions are randomly unavailable depending on the epochs. In addition to the theoretical results found, this part emphasizes the importance of focusing the exploration where it is needed.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.