3. Literature survey: Classification of sleepiness in drivers
3.3 Machine Learning Algorithms
3.3.3 Types of Machine Learning Algorithms
MLA can be classified in three different categories: supervised learning, unsupervised learning and reinforcement learning (Harrington, 2012; Bell, 2015; Marsland, 2015; Murphy, 2012; Alpaydin, 2010).
3.3.3.1 Supervised learning
In supervised learning, examples containing information of the desired output are presented to the MLA. For the MLAs presented in this chapter, every dataset that is input into a MLA is called a feature set. A feature set is the set of variables in the dataset, e.g. size of a house in a specific area, size of an engine in certain car models, the lane deviation of the driver. The feature vector, i.e. an instantiation of the feature set, is the data the MLA will use to learn and train itself. In supervised learning, the dataset also contains the expected value associated to the feature set. These values are called target set, e.g. price of a house according to specification of the house, acceleration power of a car according to specifications of the engine, level of sleepiness of a driver according to the driving behaviour. Most of the times, the feature set is not specified as a continuous value but as a class. The feature sets are defined into classes (discrete sets or categories) according to their properties or attributes. The target vector, an instantiation of the target set, contains the class that the MLA is trying to predict using the feature vectors. The MLA will try to learn by reducing the error between the predicted value by the MLA and the expected value. This can be explained in the following example.
Imagine that a person decides to collect the data of X houses around a specific area. The person obtains the size of the houses around a specific area, the number of rooms and the prices of each house. This means that the person has a dataset composed by a features set characteristics of the house and target set the price of the house. That person is interested in using a MLA that given the size of any house and the number of rooms, the MLA can predict the price of the house. For each set of feature vectors in the feature set, i.e. characteristics of the house, the MLA will
predict a price, which will be compared with the real price of the house. In each ‘learning’ iteration process, the MLA will update its parameters to reduce the error, i.e. the difference between the predicted value and the real value of the house.
3.3.3.2 Unsupervised learning
The second types of MLA are unsupervised learning algorithms. In comparison to the supervised learning algorithms, the unsupervised learning algorithms do not have a target set, i.e. the user does not know the expected output. These types of algorithms are used when the user does not have the knowledge of the real value for the features obtained. In this case, the MLA will try to find patterns in the features input by the user and create its own target class. These types of MLA are used by companies like Amazon (Linden, Smith & York, 2003). Researchers at Amazon are not sure how people can be categorised for targeted marketing (by novels genre, preferred sports, preferred videogames, etc.). Instead they allow the MLA to find clusters depending on the items people buy through the website. After finding different clusters, e.g. people who like crime novels, if a person is considered to be part of that cluster, Amazon’s MLA will recommend books that belong to the crime novel cluster.
3.3.3.3 Reinforcement learning
The final type is reinforcement learning. These type of algorithms are not concerned with a specific output from a specific set of features, instead they are concerned with a correct sequence of actions that might reach a desired goal (Alpaydin, 2010). One example for these types of algorithms can be found in machines dedicated to playing chess (Alpaydin, 2010; Block, et al., 2008; Campbell, Hoane Jr. & Hsu, 2002; Hsu, 1999). The output of a specific play (action) might not be important meanwhile the policy, i.e. sequence of correct actions, lead to the desired output, i.e. winning the game (Harrington, 2012; Bell, 2015; Marsland, 2015; Murphy, 2012; Alpaydin, 2010; Block, et al., 2008; Campbell, Hoane Jr. & Hsu, 2002).
This means that the type of MLA that should be chosen depends on the goal, the features set and the target set. In the field of driving and sleeping, many researchers have used MLAs that can predict sleepiness while driving (Yeo et al.,
2009; Shuyan & Gangtie, 2009; Yang et al., 2010; Patel et al., 2011). To train the MLA, the researchers input the behaviour of people when they are driving while being awake and their behaviour when they are driving whilst sleepy. The MLA will then obtain the behaviour of a new driver and predict if the driver is in an awake or sleepy state. This is a supervised learning approach, i.e. learning by example. In the following section, the different types of supervised learning algorithms used to predict sleepiness in driving presented in literature will be discussed.