Machine Learning – An
Introduction
In general, Machine Learning is a
type of artificial intelligence that provide machines with the ability to learn
without being explicitly programmed. In order to make the machine robust, some
amount of pre-training is required so that the machine can learn new and
similar (can be defined in different dimensions) instances of a particular task
on its own (without any human intervention).
Why Machine Learning?
Basically, for any task, the
approaches can be classified into two broad categories: Rule-based and Machine
Learning approaches.
Rule-based approaches are nothing
but writing if-then-else possible conditions to achieve the desired outcome. Framing
rules is the starting point in understanding the machine learning approaches,
because any machine learning algorithms can perform their learning through set
of features (properties or characteristics - vary for different domains and
tasks).
Rules based approaches have some
limitations and disadvantages which are overcome in machine learning
approaches. One such limitation is the rules are domain specific. It is
difficult to design the rules as domain independent. It leads to the conflict
of the existing rules, which results in overfitting and/or producing some false
positive and true negative results. It is also hard to design abstract level of
rules where the desired outcome could not be reached with the abstract rules.
Moreover, it is difficult to capture some specific classes through generic
rules.
Thus, to overcome the limitations
and barriers, the machine learning algorithms are introduced.
An Example - Rule-based
Approach Vs Machine Learning
Let us consider a chess board
which have 8X8 matrix, consists of 8 rows and 8 columns. The task is to build
the rules as well as machine learning algorithm for this game.
Figure 1
Figure 1 shows a sample chess
board. There are certain steps to frame the rules. They are:
1. In
order to frame the rules for this chess board, we need to define rules for each
chess pieces or coins with some pre conditions
2. Rules
needs to be framed for each row and column
3. Rules
are also framed as per the opponent move.
4. Some
chess pieces or coins have restrictions which needs to be defined in the rules
Based on these conditions, the
rules are framed which are big enough to write when the task is big and the
data is too large.
On the other hand, machine
learning approaches require a set of probabilities for example, to perform the
same actions done by the set of rules.
Here too, a set of preconditions
and/or probabilities are required to start with.
1. Initial
probability for all the chess pieces or coins
2. Transition
probability for each move of each coin
3. Emission
probability – the outcome of the each move. If the opponent coin is defeated,
then the probability of opponent’s chess piece is zero and the one who defeats
is some increased probability from the previous state.
Figure 2 – Moves highlighted for White Horse
From
Figure-2, it can be seen that, there are six different ways to move the White
horse to save from Black King. For rule-based approach, six different rules are
needed. Among six, four are safe and two are not. These are also captured by
the rules where more complex rules are needed to reslove this.
In
machine learning, the probabilities of each place and each coin should decide
on which place is the best to choose to move the white horse. So six different
probabilities will be computed and among them, four have some moderate
probabilities and two of them have some low proabilities. Based on these
values, the correct position will be selected to place the White Horse.
Likewise,
the probabilities for other chess pieces are also computed.
Thus machine learning is
required for any task which makes the system robust. 