Honors Theses and Capstones

Date of Award

Spring 2023

Project Type

Senior Honors Thesis

College or School



Computer Science

Program or Major

Computer Science

Degree Name

Bachelor of Science

First Advisor

Wheeler Ruml


Partially observable stochastic games (POSGs) are difficult domains to plan in because they feature multiple agents with potentially opposing goals, parts of the world are hidden from the agents, and some actions have random outcomes. It is infeasible to solve a large POSG optimally. While it may be tempting to design a specialized algorithm for finding suboptimal solutions to a particular POSG, general-purpose planning algorithms can work just as well, but with less complexity and domain knowledge required. I explore this idea in two different POSGs: Navy Defense and Duelyst.

In Navy Defense, I show that a specialized algorithm framework, goal-driven autonomy, which requires a complex subsystem separate from the planner for explicitly reasoning about goals, is unnecessary, as simple general planners such as hindsight optimization exhibit implicit goal reasoning and have strong performance.

In Duelyst, I show that a specialized expert-rule-based AI can be consistently beaten by a simple general planner using only a small amount of domain knowledge. I also introduce a modification to Monte Carlo tree search that increases performance when rollouts are slow and there are time constraints on planning.