Lesson 01 - Introduction to Machine Learning

1.1 Introduction
People have discovered knowledge for solving problems. In a broader sense, there are two
kinds of approaches as cognitive systems and machine learning to solve problems. The
cognitive system approach uses knowledge in the form of theories, models, etc. which can
be stored in computers to solve problems. Expert systems and Natural Language processing
systems are classic examples about cognitive systems. This is because they are based on
stored knowledge in knowledge bases, data bases, dictionaries, corpus, etc. We have
discussed such Cognitive systems in the module on Expert Systems and natural Language
Processing. In contrast, this module is on machine learning. The fundamental idea behind
machine learning is the use of huge collection of examples, without considering any theory
or model behind those examples, to solve problems. This module covers three major
machine learning techniques namely, decision trees, artificial neural networks and genetic
algorithms.

1.2 What is machine learning?

We have already stated that modern approach to Artificial Intelligence has been identified
as the development of Rational Agent. The rational agent is a kid of agent who does the
right thing. In simple term, just like a travel agent, an agent is a software program works
on behalf of his master. Such an agent program has many exciting features, including,
autonomy, reactivity, reactivity, adaptability, rationality, and more importantly the learning
ability. Learning has been a strategy for an Agent to improve himself. Otherwise,
performance of an Agent program remains unchanged over the time. However, just like
human agents, software agents also improve their performance over the time through a
process of learning.

It would be an interesting question to ask how an Agent learns. Before investigating how an
agent learns, let us think of how we learn. One of the most powerful learning strategy used
by humans can be seen as ‘learning from examples’. For instance, without out knowing any
theory, skill workers can still perform some time much better than persons with theoretical
knowledge. As a well known example, students prepare for GCE (A/L) and even for
university level examination just by doing a large number of sums (examples) from a
particular subject. They practice hundreds of examples, sometime, without having a proper
idea about theory behind those examples.

Although, some may not appreciate the problem solving skills that are acquired by mere
training, this approach still works very well in some cases. For example, trained persons
can solve problems much faster and more rational manner than a theoretical person. More
examples are used for the training, higher the capacity to solve problems.

Obviously, the real world is full of situations, where we do not know have any theory,
model, algorithms, formula, etc. to resolve them. For example, illegible hand writing can
be understood by humans yet, we cannot develop theory or formulae to describe hand
writing of a person. As other examples, data for weather forecasting and stock market
predictions are only collection of data without fitting into any theory, model or algorithm.
In view of the above, it would be interesting to develop computer programs that learn from
examples, without being worried about model or the algorithm that describe such systems.
The techniques that enable computers to learn from examples are called machine learning
techniques. Such machine learning techniques developed by Artificial Intelligence provides
solutions for real world problems, which could not be solved otherwise.

1.3 Learning Agent

Before going into details of how machines learn, let us discuss how we really learn.
Although people have been learning right through their lives, even educated people may
not be able to explain how they have learned. However, educationists have done research
and developed various theories to facilitate different kinds of learning. For the
completeness sake, here we present the underlying mental process in our minds when we
learn in general. However, the process stated here may not be directly applicable for
machine learning or learning from examples. The learning model stated here is a kind of
conscious learning. This does not mean that learning from examples does not require a
conscious process within the mind.

Learning model adapted from Russell and Norvig 2003 is shown in Figure 1.1. The learning
model shown in Figure 1.1 comprises of four main activities as follows.

• Critique
• Learning element
• Performance element
• Problem generator

Figure 1.1 Learning Agent

Before discussing the four components, let briefly describe the role of a Learning Agents.
As we already know, all agent act as per the input coming in. However, a learning an agent
not only acts but also improves the agent’s ability to act in the future. Such a learning
process can ranges from memorization to creation of new knowledge. Now we briefly
describe the role of the above four components of a learning agent.
Critic
Learning Agent begins with an input from the environment. Then the agent executes the
function of being critique and tells the learning element how well the agent is doing
(success or failure). In a learning process, we have to be aware of whether we learn or do
not learn some thing come to our mind. In other words, one cannot be successful in
learning, if he is not conscious of what is being learned.

Problem generator
In a successful learning session, we are always curious about what is next. Stated another
way, there will be a kind of curiosity or problem created in our minds, if we are to attach
into the learning process. Our lives are always driven by our motivations to find a solution
for a certain question. In fact, this is why we trend to ask questions to check whether a
person has learned something. On the other hand, we find that it is boring to learn
something, we already know. Therefore, sense of ignorance or a problem in mind would be
a driving force for learning.

Performance element
The performance element is the entire learning agent based on so far operations. As such
the performance element accepts an input sequence coming in and decides on actions. The
performance element includes different types of knowledge such as;
• mapping from conditions on the current state to actions
• information about the way the world evolves
• information about the possible actions the agent can take
• information about the result of possible actions by the agent

The design of performance element is closely coupled with the design of the learning
element stated below.

Learning element
The learning element determines how the performance element (the agent itself) should be
modified to do better in the future. Yet another major task of learning element is to ensure
the efficiency of the performance element. For example, agents are able to do the similar
things without going through the previous steps from the scratch. Building of such
capabilities to learning agents is a key task of learning element. A large number of methods
for implementing learning elements have been proposed by the machine learning
community.

The learning element distinguishes the learning agent from ordinary agents. Therefore,
discussion on learning element would be the key task of describing a Learning Agent. The
design of learning element is affected by many considerations such as

• Which components of the performance elements are to be improved

• Which representation is used for those components
• What feedback is available
• What priori knowledge is available
1.4 Forms of Learning
In the previous section we stated that learning element of the learning Agent modifies the
performance element so that agent makes better decisions. Nature of the feedback has
been identified as one of the most important factor in learning. Machine learning
community has identified the following three cases of learning in this sense.

• Supervised
• Unsupervised
• Reinforcement

Supervised learning
Supervised learning is about learning from inputs as well as outputs. This means, each input
is learned, by knowing what the output should be. Thus output is known before hand and it
is called desired output. This is what we exactly do when learn something by knowing what
the outcome or the objective before hand. Decision Tress and Artificial Neural Networks are
ideal for supervised learning.

Unsupervised learning
In this case, an agent learn from inputs, yet without having any idea about what the output
should be. This is very similar to learning of new subject for the first time, with no clue
about what is being learned and why? Artificial neural Networks can effectively model
unsupervised training data.

Reinforcement learning
This learning strategy is rather different. Here the learner goes ahead in a rather
unsupervised manner in learning. At different points learner will be indicated whether he is
right or wrong (reward and punishment). Learning session would be satisfactory, if the
learner has more reward. Thus the leaner learns from reinforcement. Genetic algorithm is
the best technology to improve reinforcement learning as it deals with the concept of
fitness function.

Other kinds of learning

Note also that from educational viewpoint there are also learning models such as root
learning, concept learning and inductive learning. Here the root learning and concept
learning is very much to do with remembering, and this is not the theme of machine
learning.

In contrast, inductive learning is about learning from input/output pairs, which is the
fundamental idea behind machine learning. Note that machine learning goes beyond mere
memorization, yet achieve the learning through generalization. The basic idea behind
generalization is the ability to detect or predict about previously unseen examples.
1.5 Summary
This lesson provided an introduction to machine learning. The machine learning concepts
were presented in comparison with cognitive systems. Primarily machine learning was
defined as learning from examples. The components of a learning Agent has been identified
with a special emphasis on learning element and the performance element. The lesson also
stated about different kinds of learning, namely, supervised, unsupervised and
reinforcement learning.

Activity 1.1
Conduct a survey and identify different types of learning discovered by Educationists.