2002/2003 Undergraduate Projects

Dr Joanna Bryson

E-mail: J.J.Bryson@bath.ac.uk

General Interests: Aritificial Intelligence, Natural Intelligence, Artificial Life, Software Systems

Samples of Specific Interests:Modular models of animal intelligence, the evolution of social behaviors, intelligent environments, tutoring systems, VR characters, autonomous robots, multi-agent models of political systems.


Title: Prototyping GUIs for editing and debugging hierarchical reactive plans.

Description:

Reactive plans are an excellent way to organize the behavior of complex artificial agents, such as VR characters or simulated animals. A developer provides a concise, robust plan which the agent can then apply in a large number of diverse, dynamic situations. However, visualizing, coding and debugging reactive plans can be tricky.

This project involves developing a GUI for a hierarchical reactive plan structure with 5 types: two sorts of primitives (actions and perceptions), and three sorts of aggregates (a simple sequence and two elaborations on the sequence.) A standard representation already exists: the GUI must read and write this representation, and facilitate its editing.

Ideally, the GUI will be platform independent, and will also have an API so that running agents or agent logs can be connected to it to facilitate debugging. The eventual goal is a platform independent tool to be used by researchers and students. However, for this project, the primary goal is to develop and test a number of prototype models.

The final writeup should contain both documentation for at least one of the developed prototypes, and some justification of the design, including user studies. Write-up may be directed towards either industrial or academic standards, depending on the student's career interests.

The reactive plan structure used in this project is called POSH and is one I developed myself. POSH planning is now being used in several different universities, so there is some chance that a good project would be used well after completion.

Pre-requisite knowledge:

Strong competence in at least one GUI development language, preferably one suited for rapid prototyping of platform-independent GUIs (e.g. TCL/TK, Python, possibly Java). Interest though not necessarily experience in A.I.

Indicative reading:

Example of a reactive plan for Artificial Life research: Hierarchy and Sequence vs. Full Parallelism in Reactive Action Selection Architectures .

A shorter & more technical description of these reactive plans: Modularity and Design in Reactive Intelligence

A longer & more fun description of the industrial application of such plans: Dragons, Bats & Evil Knights: A Three-Layer Design Approach to Character Based Creative Play


Title: Replicating a Model of Primate Social Order

Description:

The objective of this project is to replicate a study suggesting that the extent to which a primate society is either egalitarian or strictly hierarchical is determined simply by the level of violence typical of alteractions between group members. The original research, done by C. K. Hemelrijk, is intended to explain the differences in social order in related species of macaques.

The purpose of replication is that it provides a seperate model which can then be used for testing further an interesting hypothesis. There are two possible outcomes of a replication. The first is that the replication may fail, or require significant extensions (such as extra assumptions) beyond what was reported in the literature. Such findings are significant in themselves.

The second possible outcome is success. In the case of the Hemelrijk model, a successful outcome would be in itself significant, because this would be the first replication. Further, the new model could then be used to test for extensions. For example, if another attribute of macaque behavior (such as behavior diversity) is varied, could level of violence be explained as a dependent variable?

The replication would be done in a multi-agent framework which has already been developed in lisp. The shape of the agent's world , their size & motion characteristics would need to be altered, but not developed from scratch. The final writeup should be in the format of a scientific paper.

Pre-requisite knowledge:

Competence in lisp programming, ability to learn and modify large software systems. This project must be done on linux. Should be interested in doing scientific research, though not necessarily as a career. This project may be useful to students considering continuing on to do research work.

Indicative reading:


Title: Modelling Transverse Patterning and the effects of Hippocampal Lesioning

Description:

The objective of this project is to determine whether a model of primate transitive inference learning can be used to explain results in learning transverse patterning in both normal and lesioned animals. The transitive model is written in Lisp.

Transitive inference is inferring that, given A > B and B > C, then A must be > C. Oddly enough, monkeys (and even rats and pigeons) seem to be able make this inference if they are able to learn the initial relations (A>B, B>C) in the first place. On the other hand, if you let a monkey (or even a child) choose between A, B and C at the same time, they are quite likely to think B is the greatest.

These results were explained by Harris & McGonigle in 1994 using a simple production rule model. Instead of learning about a number of pairs, the agent seems to learn about 1) given a bunch of options, which one should I attend to? and 2) given that I'm currently looking at option A, should I take it or avoid it? I have recently tested and extended this work into a learning system using a two-tier neural-network representation. The hypothesis to be tested in this project is that the first tier (which represents the association of a stimulus to an action) corresponds to the enthorinal cortex, and the second teir (which represents priorities between the different associative rules) corresponds to the rest of the hippocampal system.

Transverse Patterning is a related problem of learning a list of pairs that cycles, e.g. A > B, B > C, and C > A. The purpose of this project is to see whether the model which explains the TI results also can explain TP results.

This project is not necessarily challenging from a programming perspective, but does require the ability to find, read and reason about neuroscience papers. This project most suits a student who wishes to go on to do research in psychology or cognitive science. The write-up should take the form of a scientific paper. Whether the results are positive or negative, a good project would probably be publishable.

Pre-requisite knowledge:

Competence in programming, ability to learn and manage software systems. Must know or learn about the scientific method, some neuroscience and hypothesis testing.

Indicative reading:


Title: Modelling Monkey Reasoning in ACT-R

Description:

The objective of this project is to evaluate a widely-used cognitive modelling package called ACT-R, with an eye to using it for teaching at Bath for undergraduates in the future. The project involves down-loading and installing the system, then writing a test system in it. The test system is a replication of existing AI models of primate transitive inference.

Transitive inference is inferring that, given A > B and B > C, then A must be > C. Oddly enough, monkeys (and even rats and pigeons) seem to be able make this inference if they are able to learn the initial relations (A>B, B>C) in the first place. On the other hand, if you let a monkey (or even a child) choose between A, B and C at the same time, they are quite likely to think B is the greatest.

These results were explained by Harris & McGonigle in 1994 using a simple production rule model. Instead of learning about a number of pairs, the agent seems to learn about 1) given a bunch of options, which one should I attend to? and 2) given that I'm currently looking at option A, should I take it or avoid it? I have recently tested and extended this work using a specialized neural-network representation. All of this code is available and in Lisp. ACT-R is also written in Lisp.

Writeup could either take the form of an industrial evaluation and system comparison on a usability basis, or a more formal academic paper comparing the two representations from a scientific modelling perspective, depending on the student's interests..

Pre-requisite knowledge:

Strong competence in Lisp is necessary. Ability and interest in reading other people's code and documentation and getting large systems running. Interest in AI and/or psychology would be beneficial. Previous knowledge of production rule systems would be useful but isn't necessary.

Indicative reading:

ACT-R

A (draft) description of my model, which also describes the Harris & McGonigle's 1994 production rule model: : Primate Errors in Transitive Inference

A very quick introduction to production rules as a part of Expert Systems


Title: Applying the Start-End Model of Human Sequence Learning to Primate Task Learning

Description:

The objective of this project is to determine whether a model developed for explaining performance in human sequence learning could improve a model of primate task learning. The primate task-learning model is written in Lisp, and currently models the learning of transitive inference. The Start-End model, developed by Richard Henson at Cambridge (PhD 1996, currently at UCL) says that sequences are represented by each element having two graded weights, one of which is higher at the beginning of the sequence and one of which is higher at the end.

Transitive inference is inferring that, given A > B and B > C, then A must be > C. Oddly enough, monkeys (and even rats and pigeons) seem to be able make this inference if they are able to learn the initial relations (A>B, B>C) in the first place. On the other hand, if you let a monkey (or even a child) choose between A, B and C at the same time, they are quite likely to think B is the greatest.

These results were explained by Harris & McGonigle in 1994 using a simple production rule model. Instead of learning about a number of pairs, the agent seems to learn about 1) given a bunch of options, which one should I attend to? and 2) given that I'm currently looking at option A, should I take it or avoid it? I have recently tested and extended this work into a learning system using a simple neural-network representation for sequences.

The point of this project is to see whether using the Start-End model for the sequence learning improves the model by making the distribution of solutions closer to that of real subjects (see Table 2 in "Primate Errors in Transitive Inference" below). This will also involve processing more data from real subjects in order to make the results clearer.

This project is not necessarily technically challenging (although a good programmer could probably extend the simulation in interesting ways.) This project most suits a student who wishes to go on to do research in psychology or cognitive science. The write-up should take the form of a scientific paper. Whether the results are positive or negative, a good project would probably be publishable.

Pre-requisite knowledge:

Competence in programming, ability to learn and manage software systems. Must know or learn about the scientific method and hypothesis testing.

Indicative reading:

A (draft) description of my model: Primate Errors in Transitive Inference

Short-term memory for serial order: the Start-End Model. Henson, R.N.A. (1998) Cognitive Psychology, 36, 73-137.

The statistics section out of any introductory psychology textbook. You might also want to look at Errors in Hypothesis Testing from Quantitative Methods I by Kenneth McGraw.


This final project wasn't originally offered, but was developed with a student who had initially chosen a project that was too difficult.

Name:     Dr. Joanna Bryson
E-mail:  J.J.Bryson@cs.bath.ac.uk
Title:   Building a tool for building a personality-based e-commerce site.

Description:

The objective of this project is to create a tool that can be used by business owners to create a web site with a sales-agent character on it.  The business owners are not expected to be programmes, so the tool's interface should be simple to use.  For example, the customers might be expected to use a pull-down menu to talk to the agent.  The business owner should be able to specify:
Standard program options might include:
The project proposal will require describing the web page the tool will build, as well as describing the tool that will do the building.  The main benefits of this project are learning to build a web page for building a web page, and learning to connect a number of existing systems (e.g. a database, a web-character agent) to a web page.

The dissertation should consist mostly of a document describing how to use the tool.  It should also include a record of how design features were chosen.  If time permits, it should also include examples of tests with both experienced programmers (e.g. other computer science students) and the target audience of non-programmers.

Pre-requisite knowledge:
Competence in web programming.  The ability to organize a system of programs.  The character agent will be from Haptek, so the project must unfortunately be written on a Windows platform, though it may eventually run on Macintosh as well.
Indicative reading: