Introduction - Design and Analysis of Learning Classifier Systems - page 2

natural language processing, medical diagnosis, and many more [19, 168]. There
is no universal method to handle all of these tasks, but a large set of different
approaches exists, each of which is specialised in particular problem classes.
Probably the most distinct differences between the numerous machine learning
methods is the type of task that they can handle, the approach that they are
designed with, and the assumptions that they are based upon. Describing firstly
a set of common machine learning task types, let us then, based on a simple
example, consider two common approaches to how one can develop machine
learning algorithms.

1.1.1 Common Machine Learning Tasks

The most common problem types of tasks that machine learning deals with are:

Supervised Learning. In such tasks a set of input/output pairs are available,
and the function between the inputs and the associated outputs is to be
learned. Given a new input, the learned relation can be used to predict
the corresponding output. An example for a supervised learning task is a
classification task: given several examples of a set of object properties and the
type of this object, a supervised learning approach can be taken to find the
relation between the properties and the associated type, which subsequently
allows us to predict the object type for a set of properties.
Unsupervised Learning. Unsupervised learning is similar to supervised learning,
with the difference that no outputs are available. Thus, rather than learning
the relationship between inputs and associated outputs, the learner builds
a model of the inputs. Consider a clustering task where several examples of
the properties of some object are given and we want to group the objects
by the similarity of their properties: this is an unsupervised learning task
because the given examples only contain the object properties, but not the
group assignment of these objects.
Sequential Decision Tasks. Such tasks are characterised by a set of states, and
a set of actions that can be performed in these states, causing a transition
to another state. The transitions are mediated by a scalar reward and the
aim of the learner is to find the action for each state that maximises the
reward in the long run. An example for such a task is in a labyrinth to find
the shortest path the goal by assigning each step (that is, each transition) a
reward of -1. As the aim is to maximise the reward, the number of steps is
minimised. The most common approach to sequential decision tasks is that of
dynamic programming and reinforcement learning: to learn the optimal value
of a state, which is the expected sum of rewards when always performing
the optimal actions from that state, and subsequently to derive the optimal
actions from these values.

There exists a wide range of different machine learning methods that deal with
each of the problem types. As we are interested in their design, let us consider
two possible design approaches to an unsupervised learning task.