Abstract
In this research, we present our computational model which performs four part har-monisation in the style of J.S. Bach. Harmonising Bach chorales is a hard AI problem, comparable to natural language understanding. In our approach, we explore the issue of gaining control in an explicit way for the chorale harmonisation tasks. Generally, the control over the search space may be from both domain dependent and domain inde-pendent control knowledge. Our explicit control emphasises domain dependent control knowledge. The control gained from domain d ependent control enables us to map a clearer relationship between the control applied and its effects. Two examples of do-main dependent control are a plan of tasks to be done and heuristics stating properties of the domain. Examples of domain independent control are notions such as temperature values in an annealing method; mutation rates in Genetic Algorithms; and weights in Artificial Neural Networks.
The appeal of the knowledge based approach lies in the accessibility to the control if required. Our system exploits this concept extensively. Control is explicitly expressed by weaving different atomic definitions {i.e. the rules, tests and measures) together with appropriate control primitives. Each expression constructed is called a control definition, which is hierarchical by nature.
One drawback of the knowledge based approach is that, as the system grows bigger, the exploitation of the new added knowledge grows exponentially. This leads to an intractable search space. To reduce this intractability problem, we partially search the search space at the meta-level. This meta-level architecture reduces the complexity in the search space by exploiting search at the meta-level which has a smaller search space.
The experiment shows that an explicitly structured control offers a greater flexibility in controlling the search space as it allows the control definitions to be manipulated and modified with great flexibility. This is a crucial clement in performing partial search over a big search space. As the control is allowed to be examined, the system also potentially supports elaborate explanations of the system activities and reflections at the meta-level.
