Information Technology Reference
In-Depth Information
of weight decay in neural networks as Gaussian priors on their weights (see
Ex. 3.4). The significant performance increase of reinforcement learning through
intelligent exploration can almost exclusively be attributed to advances in their
theoretical understanding [126, 28, 204]. Correspondingly, while further impro-
vement of the already competitive performance of LCS in supervised learning
tasks cannot be guaranteed through advances from the theoretical side, such
advances unquestionably increase their understanding and provide a different
perspective.
Of course, the presented methodology is by no means supposed to be the
ultimate and only approach to design LCS. It is not the aim to stifle the inno-
vation in this field. Rather, its uptake is promoted for well-defined tasks such
as regression and classification tasks, due to the obvious advantages that this
approach promises. Also, given that Sutton's value-function hypothesis [206] is
correct, and value function learning is the only ecient way to handle sequential
decision tasks, then these tasks are most likely best approached by taking the
model-centred view as well. On the other hand, given that the task does not
fall into these categories (for example, [197]), then an ad-hoc approach without
strong formal foundations might still be the preferred choice for designing LCS.
However, even following the outlined route leaves significant space for design
variations in how to formulate the model, and in particular which method to
develop or apply to search the space of possible model structures.
Overall, with the presented perspective, the answer to “What is a Learning
Classifier System?” is: a family of models that are defined by a global model
being formed by a set of localised models known as classifiers, an approach for
comparing competing models with respect to their suitability in representing the
data, and a method to search the space of sets of classifiers to provide a good
model for the problem at hand. Thus, the model was added to the method.
