Information Technology Reference
In-Depth Information
3 A Learning Classifier Systems Model
Specifying the model that is formed by a set of classifiers is central to the model-
based approach. On one hand it explicitly defines the assumptions that are made
about the problem that we want to solve, and on the other hand it determines
the training methods that can be used to provide a solution. This chapter gives
a conceptual overview over the LCS model, which is turned into a probabilistic
formulation in the next chapter.
As specified in Chap. 1, the tasks that LCS are commonly applied to are re-
gression tasks, classification tasks, and sequential decision tasks. The underlying
theme of providing solutions to these tasks is to build a model that maps a set
of observed inputs to their associated outputs. Taking the generative view, we
assume that the observed input/output pairs are the result of a possibly stocha-
stic process that generates an output for each associated input. Thus, the role
of the model is to provide a good representation of the data-generating process.
As the data-generating process is not directly accessible, the number of availa-
ble observations is generally finite, and the observations themselves possibly
noisy, the process properties need to be induced from these finite observations.
Therefore, we are required to make assumptions about the nature of this process
which are expressed through the model that is assumed.
Staying close to the LCS philosophy, this model is given by a set of localised
models that are combined to a global model. In LCS terms the localised models
are the classifiers with their localisation being determined by which inputs they
match, and the global model is determined by how the classifier predictions
are combined to provide a global prediction. Acquiring such a model structure
has several consequences on how it is trained, the most significant being that
it is conceptually separable into a two-step procedure: firstly, we want to find
a good number of classifiers and their localisation, and secondly we want to
train this set of classifiers to be a seemingly good representation of the data-
generation process. Both steps are closely interlinked and need to be dealt with
in combination.
A more detailed definition of the tasks and the general concept of modelling
the data-generating process is given in Sect. 3.1, after which Sect. 3.2 introduces
 
Search WWH ::




Custom Search