Code

COMPGC26 (Also taught as: COMP3058)

Year

MSc

Prerequisites

A strong background in university-level maths (in particular logic)

Term

2

Taught By

Denise Gorse (50%)
Anthony Hunter (50%)

Aims

This course introduces artificial intelligence and neural computing as both technical subjects and as fields of intellectual activity. The overall targets are: (1) to present basic methods of expressing knowledge in forms suitable for holding in computing systems, together with methods for deriving consequences from that knowledge by automated reasoning; (2) to present basic methods for learning knowledge; and (3) to introduce neural computing as an alternative knowledge acquisition/representation paradigm, to explain its basic principles and their relationship to neurobiological models, to describe a range of neural computing techniques and their application areas.

Learning Outcomes

Ability to identify problems that can be expressed in terms of search problems or logic problems, and translate them into the appropriate form, and know how they could be addressed using an algorithmic approach. Ability to identify problems that can be expressed in terms of neural networks, and to select an appropriate learning methodology for the problem area.

Scope of the Subject

Nature and goals of AI
Application areas

Searching state-spaces

Use of states and transitions to model problems
Breadth-first, depth-first and related types of search
A* search algorithm
Use of heuristics in search

Reasoning in logic

Brief revision of propositional and predicate logic
Different characterisations of reasoning
Generalized modus ponens
Resolution
Forward and backward chaining

Knowledge Representation

Diversity of knowledge
Inheritance hierarchies
Semantic networks
Knowledgebase ontologies

Handling uncertainty

Diversity of uncertainty
Inconsistency
Dempster-Shafer theory

Machine Learning

Induction of knowledge
Decision tree learning algorithms

Intelligent agents

An architecture for intelligent agents
Argumentation
Decision-making

Nature and Goals of Neural Computing

Comparison with rule-based AI
Overview of network architectures and learning paradigms

Binary Decision Neurons

The McCullough-Pitts model
Single-layer perceptrons and their limitations

The Multilayer Perceptron

The sigmoid output function
Hidden units and feature detectors
Training by error backpropagation
The error surface and local minima
Generalisation, how to avoid 'overtraining'

The Hopfield Model

Content addressable memories and attractor nets
Hopfield energy function
Setting the weights
Storage capacity

Self-Organising Nets

Topographic maps in the brain
The Kohonen self-organising feature map

Lecture presentations.

The course has the following assessment components:

• Written Examination (2.5 hours, 90%)
• Coursework Section (2 pieces, 10%)

To pass this course, students must:

• Obtain an overall pass mark of 50% for all sections combined

The examination rubric is:
Answer three from six questions set, at least one from

each of section A (AI) and section B (neural computing). All questions carry equal marks.

Artificial Intelligence - A Modern Approach; First Edition; Prentice Hall; ISBN: 0-13-103805-2

[Background reading] Neural Computing: An Introduction; R Beale and T Jackson;

Institute of Physics Publishing; ISBN: 0-85-274262-2