Nov 19, 2018
Jolley Hall, Room 309
"Stochastic Goal Recognition Design (S-GRD)"
Adviser: William Yeoh
Goal recognition is the problem of identifying the goal of an agent through the observation of its actions. In 2014 researchers introduced and formulated a new related problem called Goal Recognition Design (GRD), which involves identifying the best ways to design or re-design the environment that agents operate in with the objective to facilitate goal recognition. The vanilla GRD model is defined under three assumptions: (1) agents in the system act optimally; (2) the agent action outcomes are deterministic; and (3) the environment is fully observable.
This dissertation proposal aims to extend the GRD model by relaxing all three assumptions, therefore, allowing a better representation of the uncertainty inherent to the physical world. We start by relaxing the second assumption, specifically, we propose the Stochastic GRD (S-GRD) problem that assumes stochastic agent action outcomes. Later, we relax the third assumption by introducing partial observability to the model: POS-GRD, and finally, we aim to further extend the model by considering suboptimal agents.
Typically, a GRD problem has two components: (1) a measure of the efficacy of goal recognition and (2) a model of possible design changes that can be applied to the environment with the objective to improve the efficacy of the original goal recognition problem. Besides modifying the original metric for each case, we have proposed a new metric that proved to be useful in certain instances where the original metric is unable to improve the model; a new type of design technique for POS-GRD problems has also been proposed.
The dissertation will conclude with an analysis of the implications of each possible combination of relaxed assumptions applied to the model.