IJsselmuiden, J. (2006) Qualitative Decision Theory and Graph Rewriting in an Adaptive Diary Assistant. Master's Thesis / Essay, Artificial Intelligence.
|
Text
JIJsselmuiden.CV.pdf - Published Version Download (4MB) | Preview |
Abstract
Developing a multiagent system (MAS) is difficult due to numerous issues. We aim to define and discuss some of these. The project contains most of the typical phases of MAS development, from literature study to implementation (testing is beyond the scope of the project). As a case-study for the development of a multiagent system, we use an adaptive diary assistant that is run by a team of autonomous agents. Creating a useful application is only a secondary goal. It is the entire process from literature study to implementation that we are interested in, not the end product. The two main techniques used for the development of this diary assistant are qualitative decision theory and graph rewriting. Qualitative decision theory provides us with a convenient architecture for the design, whereas graph rewriting is used for implementation. Besides defining and discussing the issues encountered during MAS development, we also explore and expand the possibilities of graph rewriting as a programming language for MAS. Functionality for qualitative decision theory and complex data manipulation is added to the graphical agent programming environment OutOfBrain. The project has many goals, but the focus lies with two simple research questions: "How can Boutilier's QDT architecture be extended to allow for temporal reasoning?" and "Can Boutilier's QDT architecture, extended with a temporal modality, be integrated into OutOfBrain?". QDT can possibly be extended with temporal reasoning by combining it with BDI's apparatus for handling time series and events. We call this combined architecture QDT+ from now on. It has three modalities, one for linear preference orderings, one for linear normality orderings and one for time trees that branch into the future and are linear in the past. This allows agents to reason about time as well as mental attitudes. We claim that such a multimodal architecture is necessary for general purpose learning. The design we end up with leads to the following conclusion. QDT can be extended to allow for temporal reasoning by combining it with certain parts of BDI. We provide the first steps toward such an architecture. The second main goal of this project is to implement QDT+ as an integral part of OutOfBrain. In this endeavour we were only partly successful. The original QDT architecture, using preferences and normalities, was indeed implemented and it is now part of OutOfBrain. However, the functionality for handling time and events was not. This is partly due to time constraints on the project. Despite this deficit, we are tempted to answer our second research question in a positive way: QDT+ can indeed be implemented to become an integrated part of OutOfBrain. Although we did not succeed in doing so ourselves, we developed a clear picture of how one would go about implementing such an architecture. The project has four secondary goals. They should be viewed as the context for our two main goals. Also, we use them to solidify the more abstract challenges of analysing issues in MAS and exploring and extending OutOfBrain. The first of these secondary goals is to create a useful application using QDT and OutOfBrain. Requirements on new software are extremely high nowadays and the adaptive diary assistant does not live up to them. Security issues for example are not dealt with. One's diary should be private and not accessible to other people. Achieving this is beyond the scope of this project. Besides this issue, many others have to be dealt with before the application could be called seaworthy. However, we decided that the diary assistant should never leave the computer-lab. It served its purpose, which was to be a test-bed to explore and extend QDT and OutOfBrain. Secondary goal number two is to successfully incorporate agent communication in order to improve the application's performance and to further our understanding of sociality in MAS. The communication standard we use, FIPA-ACL, proved to be highly intuitive and perfectly capable of modelling fairly complex communication flows. Therefore, we encourage other researchers and engineers to make use of it. The third context-goal is to develop a powerful learning method to facilitate user modelling. Again, we aim to improve the diary assistant with that, but we also desire to expand our own knowledge of the topic. We were only partly successful in this endeavour. The learning technique used in the application is not general purpose; it can only operate in a highly constrained, predefined problem space that fails to impress. The fourth and final secondary goal is to use standards from cognitive ergonomics to improve the interface of the application. We hardly pursued this goal in practice, but we did end up with a usable interface thanks to our own insights and Rockingstone's experience with developing database applications. Because of our research goals, we choose to make use of a team of agents as opposed to a single agent. From an engineering point of view, it would be better to use a single agent, because the system is not distributed in space and henceforth not a true multiagent system. In the current version of the programming environment, the world, including the agents, is represented in a single graph. We simulate distributedness by posing constraints on the agents' scope of access and control over the environment. They need to work together, since each member has unique capabilities. To successfully implement the BDI components as well, one would have to develop a way to store the history of the world. This would result in not one, but a whole series of OutOfBrain graphs, one for each time step. Each one is a world that can contain anything from first order sentences to complex statements, combining temporal operators with mental attitudes. This is something for potential future work. Another possible topic for future studies is the general purpose learning method we mentioned. It would require a module that extracts logical formulas from raw input, coming from the real world, an experimental setting, a simulation or a database. Needless to say, this is an unsolved problem and many people are working on it. A learning method for QDT+ would have to transform histories of percepts to formulas containing normalities and preferences, combined with temporal operators. We believe that MAS will play a central role in the future world of diary-keeping. The best approach is to have a single agent represent his user instead of having an entire team of agents per user. We envision a company where everybody caries a PDA that acts as a personal secretary. After a single button-press by the chairman of the board, a meeting can be scheduled that fits everybody's appointments and preferences as much as possible. The personal secretaries negotiate and form alliances in order to satisfy (or rather satisfice) the desires of their masters.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 15 Feb 2018 07:30 |
| Last Modified: | 15 Feb 2018 07:30 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/9070 |
Actions (login required)
 |
View Item |