Mechanical & Industrial Engrg 624 - MachLrning/DynamDecisionMaking

Dynamic or sequential decision-making tackles a type of problems where decisions need to change over time to adapt to changing environments, occurring as a consequence of natural environmental dynamics and/or influence of prior decisions. These type of problems are encountered in a variety of fields. Examples include intervention decisions for epidemic control, personalized health decisions, production and inventory planning under dynamic demand and supplies, clean energy transitioning for sustainable development, traffic light control, cyber-physical systems control, games, and natural language generation. We will study suitable algorithms for optimization of such dynamic decisions, typically discussed under the domain of control optimization. Topics covered include neural networks, Markov decision processes, dynamic programming, and reinforcement learning. Training of machine learning algorithms rely on large data, which in some settings are unavailable or infeasible to collect. In such cases, simulation serves as a useful environment for generation of data and training of machine learning algorithms, which are then collectively referred to as simulation-based optimization. This class will build on concepts from Markov chain, simulation, and optimization, and thus provide hands-on experience in integrating knowledge from prior classes. Assignments in the 600-level class will have a methodological component that requires a deeper understanding of the techniques studied, beyond their computational implementation.

Prerequisites are coding in Python, simulation (MIE 373), and Markov chains and processes (MIE 380).