PhD Thesis Defense: Shuwa Miura, Optimized Resource Allocation for Serving Deep Learning Models
Content
Speaker
Abstract
As autonomous agents increasingly share spaces with people, it is crucial that these agents are aware of how their behaviors are interpreted. Previous research has explored various methods for communicating or concealing agents'
goals, intentions, and capabilities through their behaviors. This thesis introduces a unifying model for generating behaviors that not only achieve desired goals but also account for how these behaviors are perceived.
The proposed model, called the Observer-Aware Markov Decision Process (OAMDP), assumes that observers interpret the agent's actions to form beliefs about its potential desires, goals, and intentions. Planning with OAMDP produces behaviors that shape desirable beliefs in observers. While OAMDP is an expressive framework capable of generating diverse observer-aware behaviors, reasoning about an observer's beliefs introduces a dependence on action histories, making OAMDPs computationally intractable in the worst case.
To address this challenge, the thesis explores several approximation algorithms for solving OAMDPs. Additionally, it proposes a variation of OAMDP with constraints on task efficiency to address the potential conflict between interpretability and task efficiency.
Advisor
Shlomo Zilberstein