Events | Mechanical Engineering

Yohan John PhD Defense

November 6, 2026 1:00 PM
End time: 2:30 PM
Yohan John Headshot
Location
Elings 1601
Type
General Event

This dissertation studies two decision-making problems in which randomness plays a central role. In robotic surveillance, randomness is a tool: a patroller facing an intelligent adversary benefits from unpredictable movement, so we design Markov chains that encode stochastic patrol routes. In inventory control, randomness is a challenge: ordering decisions must balance costs against uncertain future demand, so we design policies and analyze their performance under stochastic dynamics.

Chapter 1 models the patrol environment as a graph and the patrol strategy as a Markov chain on that graph. We first extend a Stackelberg game formulation to heterogeneous defenses across the environment, derive efficiently computable patrol strategies with bounded suboptimality for several graph classes, and co-optimize defensive resources and patrol strategy. We then present RoSSO, a Python package that uses JAX to optimize single- and multi-robot patrol strategies on general graphs, and apply it to a San Francisco police district. Finally, we propose Markov chain lifting, in which several chain states may correspond to the same graph node. We prove bounds on the improvement of lifted over non-lifted chains, show that lifting can deceive an adversary who observes only the patroller's location on the graph, and provide a tractable optimization method that performs well on sparse real-world graphs.

Chapter 2 casts industrial refrigeration as inventory control: the refrigerated product stores cooling as inventory, heat loads act as demand, and compressor energy acts as the ordering cost. From a first-principles thermodynamic model, we show that compressor work is convex in heat removal. For the resulting single-location problem with convex ordering costs, we bound the savings of load shifting over a myopic policy and propose a tractable heuristic with bounded suboptimality. We then model facilities in which multiple rooms share compressors as a multi-location problem coupled through a nonlinear ordering cost, and bound the suboptimality of decoupled policies in terms of that nonlinearity.