**5. Algorithmic foundations**

This section is devoted to describe two important figures in pattern recognition: *backpropagation neural networks* BPNN and *genetic algorithms* GA. The BPNN is used as a reference classifier to compare the performance of the approach presented here to the word recognition problem. The GA is an integral part of the generation of the features in the proposed technique.

### **5.1 Learning paradigms**

There are several major paradigms, or approaches, to machine learning. These include supervised, unsupervised, and reinforcement learning. In addition, many researchers and application developers combine two o more of these learning approaches into one system [23].

*Supervised learning* is the most common form of learning and is sometimes called programming by example. The learning system is trained by showing it examples of the problem state or attributes along with the desired output or action. The learning system make a prediction based on the inputs and if the output differs from the desired output, then the system is adjusted or adapted to produce the correct output. This process is repeated over and over until the system learns to make accurate classifications or predictions. Historical data from databases, sensor logs, or trace logs is often used as the training or example data.

*Unsupervised learning* is used when the learning system needs to recognize similarities between inputs or to identify features in the input data. The data is presented to the system, and it adapts so that it partitions the data into groups. The clustering or segmenting process continues until the system places the same data into the same group on successive passes over the data. An unsupervised learning algorithm performs a type of feature detection where important common attributes in the data are extracted.

*Reinforcement learning* is a type of supervised learning used when explicit input/output pairs of training data are not available. It can be used in cases where there is a sequence of inputs and the desired output is only known after the specific sequence occurs. This process of identifying the relationship between a series of input values and a later output value is called temporal credit assignment. Because we provide less specific error information, reinforcement learning usually takes longer than supervised learning and is less efficient. However, in many situations, having exact prior information about the desired outcome is not possible. In many ways, reinforcement learning is the most realistic form of learning.

Another important distinction in learning systems is whether the learning is done on-line or off-line. On-line learning means that the system is sent out to perform its tasks and that it can learn or adapt after each transaction is processed. On-line learning is like on the job training and places severe requirements on the learning algorithms. It must be very fast and very stable. Off-line learning, on the other hand, is more like a business seminar. You take your salespeople off the floor and place them in an environment where they can focus on improving their skills without distractions. After a suitable training period, they are sent out to apply their new found knowledge and skills. In an intelligent system context, this means that we would gather data from situations that the systems have experienced. We could then augment this data with information about the desired system response to build a training data set. Once we have this database we can use it to modify the behavior of our system.
