AbstractThis thesis presents a new computational learning algorithm that is capable of learning new concepts for the game of Go from both textual descriptions and pictorial examples. The 'GOaL' (GO and Learn) computational model uses this algorithm to learn descriptions of concepts that it studies. These descriptions enable the model to recognise a variety of positions by allowing the recognition of salient features of concepts. Experimental results are included that are derived from work on computational learning in cognitive science. A comparison is made between machine learning done by the model and learning done by people. It is shown that the 'GOaL' model, when learning and completing performance tasks, behaves in similar way to that of the people that were studied. This thesis answers the research question: can a computational model learn new concepts about the game of Go, at different levels of expertise, from lessons containing both textual descriptions of concepts together with pictorial examples and some simple facts about the board geometry.
This work provides an original contribution to knowledge in the sphere of machine learning, in the form of a computational model. This work also assists in the understanding of how people learn concepts for the game of Go. The contribution to knowledge is significant in that it shows that it is possible for a computational model to learn not only from pictorial examples but also from textual descriptions. People use complex, conceptual knowledge to plan and reason and, in studying and playing the game of Go, they acquire general principles of play from specific cases. An important aspect of the model is that it demonstrates that it is possible to learn representations that are defined in terms of abstract features that may be used in the identification of spatially varied instances of concepts.
|Date of Award||2001|