Software agents & human behavior

Danial, Syed Nasir (2019) Software agents & human behavior. Doctoral (PhD) thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (5MB)

Abstract

People make important decisions in emergencies. Often these decisions involve high stakes in terms of lives and property. Bhopal disaster (1984), Piper Alpha disaster (1988), Montara blowout (2009), and explosion on Deepwater Horizon (2010) are a few examples among many industrial incidents. In these incidents, those who were in-charge took critical decisions under various ental stressors such as time, fatigue, and panic. This thesis presents an application of naturalistic decision-making (NDM), which is a recent decision-making theory inspired by experts making decisions in real emergencies. This study develops an intelligent agent model that can be programed to make human-like decisions in emergencies. The agent model has three major components: (1) A spatial learning module, which the agent uses to learn escape routes that are designated routes in a facility for emergency evacuation, (2) a situation recognition module, which is used to recognize or distinguish among evolving emergency situations, and (3) a decision-support module, which exploits modules in (1) and (2), and implements an NDM based decision-logic for producing human-like decisions in emergencies. The spatial learning module comprises a generalized stochastic Petri net-based model of spatial learning. The model classifies routes into five classes based on landmarks, which are objects with salient spatial features. These classes deal with the question of how difficult a landmark turns out to be when an agent observes it the first time during a route traversal. An extension to the spatial learning model is also proposed where the question of how successive route traversals may impact retention of a route in the agent’s memory is investigated. The situation awareness module uses Markov logic network (MLN) to define different offshore emergency situations using First-order Logic (FOL) rules. The purpose of this module is to give the agent the necessary experience of dealing with emergencies. The potential of this module lies in the fact that different training samples can be used to produce agents having different experience or capability to deal with an emergency situation. To demonstrate this fact, two agents were developed and trained using two different sets of empirical observations. The two are found to be different in recognizing the prepare-to-abandon-platform alarm (PAPA ), and similar to each other in recognition of an emergency using other cues. Finally, the decision-support module is proposed as a union of spatial-learning module, situation awareness module, and NDM based decision-logic. The NDM-based decision-logic is inspired by Klein’s (1998) recognition primed decision-making (RPDM) model. The agent’s attitudes related to decision-making as per the RPDM are represented in the form of belief, desire, and intention (BDI). The decision-logic involves recognition of situations based on experience (as proposed in situation-recognition module), and recognition of situations based on classification, where ontological classification is used to guide the agent in cases where the agent’s experience about confronting a situation is inadequate. At the planning stage, the decision-logic exploits the agent’s spatial knowledge (as proposed in spatial-learning module) about the layout of the environment to make adjustments in the course of actions relevant to a decision that has already been made as a by-product of situation recognition. The proposed agent model has potential to be used to improve virtual training environment’s fidelity by adding agents that exhibit human-like intelligence in performing tasks related to emergency evacuation. Notwithstanding, the potential to exploit the basis provided here, in the form of an agent representing human fallibility, should not be ignored for fields like human reliability analysis.

Actions (login required)

View Item