Modelling human behaviour with BDI agents

Modelling Human Behaviour with BDI Agents by Emma Jane Norling Chairperson of Supervisory Committee: Professor Liz Sonenberg Although the BDI framework was not designed for human modelling applications, it has been used with considerable success in this area. The work presented here examines some of these applications to identify the strengths and weaknesses of the use of BDI-based frameworks for this purpose, and demonstrates how these weaknesses can be addressed while preserving the strengths. The key strength that is identified is the framework’s folk-psychological roots, which facilitate the knowledge acquisition and representation process when building models. Unsurprisingly, because the framework was not designed for this purpose, several shortcomings are also identified. These fall into three different classes. Firstly, although the folk-psychological roots mean that the framework captures a human-like reasoning process, it is at a very abstract level. There are many generic aspects of human behaviour – things that are common to all people across all tasks – which are not captured in the framework. If a modeller wishes to take these things into account in a model, they must explicitly encode them, replicating this effort for every model. To reduce modellers’ workload and increase consistency, it is desirable to incorporate such features into the framework. Secondly, although the folk-psychological roots facilitate knowledge acquisition, there is no standardised approach to this process, and without experience it can be very difficult to gather the appropriate knowledge from the subjects to design and build models. And finally, these models must interface with external environments in which they ‘exist.’ There are often mismatches in the data representation level which hinder this process. This work makes contributions to dealing with each of these problems, drawing largely on the folk-psychological roots that underpin the framework. The major contribution is to present a systematic approach to extending the BDI framework to incorporate further generic aspects of human behaviour and to demonstrate this approach with two different extensions. A further contribution is to present a knowledge acquisition methodology which gives modellers a structured approach to this process. The problems at the agent-environment interface are not straightforward to solve, because sometimes the problem lies in the way that the environment accepts and receives data. Rather than offering the golden solution to this problem, the contribution provided here is to highlight the different types of mismatches that may occur, so that modellers may recognise them early and adapt their approach to accommodate them. This is to certify that (i) the thesis comprises only my original work, (ii) due acknowledgement has been made in the text to all other material used, (iii) the thesis is less than 100,000 words in length, exclusive of table, maps, bibliographies, appendices and footnotes.