One way to implement the first two kinds of learning is to specify a model of the coupling or the interaction of parameters and results, and from that model derive an appropriate learning mechanism that will find a parametrisation for that model that will enable good performance to be obtained. From this starting point it has been possible to show how equal, or better performance can be obtained by using iearning mechanisms which are neither derived from nor require a model of the task being learned. Instead, by combining iteration and a task specific profit function it is possible to use a generic behavioural module based on a learning mechanism to achieve the task.

Iteration and a task specific profit function can also be used to learn which behavioural module from a pool of equally competent modules is the best at any one time to use to achieve a particular goal. Like the other two kinds of learning, this successfully automates an otherwise difficult test and evaluation process that would have to be performed by a developer. In doing so effectively, it, like the other learning that has been used here, shows that instead of being a peripheral issue to be introduced to a working system, learning, carried out in the right way, can be instrumental in the production of that working system.

The research described in this thesis examines how machine learning mechanisms can be used in an assembly robot system to improve the reliability of the system and reduce the development workload, without reducing the flexibility of the system. The justification foi' this is that for a robot to be performing effectively it is frequently necessary to have gained experience of its performance under a particular configuration before that configuration can be altered to produce a performance improvement. Machine learning mechanisms can automate this activity of testing, evaluating and then changing.

• The selection of the optimal parameter settings. • The establishment of the action-sensor couplings which are necessary for the effective handling of uncertainty. • Choosing which way to achieve a goal.

From studying how other researchers have developed working robot systems the activities which require most effort and experimentation are:-