The reinforcement learning etc.)6.2.1 Efficacy & GeneralisationAThe reinforcement learning etc.)6.2.1 Efficacy & GeneralisationA

The results discovered demonstrate the proposed approaches superior performance to the Chow-Liu algorithm when applied to a narrow test case. However, this superior performance decreasesas the number of dimensions within the target BN increases, likely a reflection of the increasedcomplexity of the task. Furthermore, this superior performance has only be demonstratedon synthetic data that has been produced in precisely the same fashion as the data that theneural network was trained on. In of itself this is not a cause for concern, but the likelihoodis that generalised performance could not be achieved without substantially increasing anddiversifying the training data corpus. Nevertheless, when considered holistically, the resultscan be considered to have broken new ground in a promising new application of CNNs.6.2 Further WorkA number of extensions are available to extend the exploratory work described within thisthesis. So, while any optimism regarding the initial findings must be tempered, clear avenuesfor improvement exist; these have been loosely grouped around the following two themes:•Efficacy Generalisation : that is improving the predictive outcomes on the existing syn-thetic problem set and extending the problem set to include other distributions and graphstructures•Innovation Integration : using the CNN model as an input into other algorithms (e.g.existing heuristic approaches, reinforcement learning etc.)6.2.1 Efficacy & GeneralisationA key predicate of this approach is the quality of the training data that the CNN is exposed to.As in all machine learning endeavours, results are likely to be underwhelming when applied todata that is significantly heterogeneous from the training data the algorithm was exposed to.While this is moderately discouraging, it should also be considered an avenue for improvement.Several approaches to expanding and diversifying the training data set are proposed below:•Synthetic Data Representativeness : Perform profiling and clustering on UCI (or other)datasets to classify/cluster the distribution types and mixing parameters;•Synthetic Data Variety : Increase the variety of data by varying the degree of orthogo-nality, the number of distribution types etc.;•Data augmentation : using techniques borrowed from image processing to expand anddiversify the existing labelled image datasets;