PhD defence of Sebastian Pilarski – Artificial Intelligence Driven Decision-Making Under Uncertainty

Abstract

Decision-making is a fundamental problem in the modern world. Technology has developed to a level where automated decision-making is used even in safety-critical systems such as self-driving cars and industrial gas turbines. The design and operation of such systems often requires decision-making without full-knowledge or information, i.e., decision-making under uncertainty.

Uncertainty can manifest itself in many ways. Notable examples include non-reliable and/or delayed information. This thesis investigates methods to advance and improve the reliability of artificial intelligence led decision-making systems under these forms of uncertainty. This thesis covers both sequential decision-making and predictive systems. These are investigated on the backdrop of two industrial application spaces: food retail and cyber-physical systems.

First, this thesis develops novel algorithms for multi-armed bandits and sequential decision-making. We present the first practical computation and indexing method for the optimal policy for Bernoulli multi-armed bandits, which was previously considered intractable for several decades. Furthermore, we provide the optimal policy for Bernoulli bandits with delayed decision outcomes. We benchmark and gauge existing popular algorithms and showcase how performance deteriorates significantly in the presence of delay. We then generalize the concepts to non-Bernoulli distributions with delay.

To exploit sequential decision-making in a practical application, we build a simulator to serve as a sandbox for experimentation to reduce food waste in food retail. We present a flexible framework capable of simulating a variety of food retail entities and their interactions. Each entity is controllable by reinforcement learning agents. We demonstrate how combining simulation with reinforcement learning can effectively reduce food waste and increase profits relative to a baseline.

Finally, the thesis investigates and provides methodologies for building more robust predictive systems in the presence of information uncertainty. Many industrial problems require decision-making with limited information or potentially unreliable information. In collaboration with Siemens Energy as industrial partner, we develop machine learning predictors used for designing aeroderivative gas turbines as complex cyber-physical systems. We also provide a methodology for deploying such machine learning predictors to existing resource-constrained control hardware.

In conclusion, this thesis provides novel decision-making techniques for various forms of uncertainty by exploiting both theoretical and practical results across different application domains.