Machine Learning (ML) is
an area in Artificial Intelligence (AI) that deals with replicating
learning processes used by humans in the form of algorithms. Ever since
Alan Turing introduced the Turing Test in 1950, a whole new sector of
reseeach emerged on creating artificial intelligence. But moving onto
ML, there are many sub-kinds/sub-topics within ML.
But what is a Learning Problem? A Learning problem can be formalized as improving the agent experience E over task T with a performance measure P. For example, to learn to play chess is a learning problem with the task (T) of playing chess. The goal for a learning is to come up with the best hypotheses (some kind of algorithm or a simple function) that when used, can perform the task T.
Now that we know what a learning problem is, let's dive into the types of Machine Learning. Most kinds of learning can be divided into 3 sub-groups: Supervised Learning, Unsupervised Learning, Reinforcement Learning. Supervised Learning uses labeled and structured data sets for training whereas unsupervised learning does not. Labeled data sets have input and output variables defined. Reinforcement learning does not use any training data sets at all.
Supervised Learning
Even within supervised learning, there are two kinds of learning: Inductive and Analytical learning.
Inductive Learning:
Analytical Learning:
Unsupervised Learning
Unsupervised Learning doesn't use labeled data sets for training the task performing hypothesis. The most famous example for this group is Clustering.
Reinforcement Learning
Reinforcement Learning does not use any training data set. Instead it used the reward-penalty concept in psychology. Given there is an agent in a certain states with certain number of actions available for him to perform, which action should it choose? The agent should choose a sequence of actions that maximize the cumulative reward at the end.
But what is a Learning Problem? A Learning problem can be formalized as improving the agent experience E over task T with a performance measure P. For example, to learn to play chess is a learning problem with the task (T) of playing chess. The goal for a learning is to come up with the best hypotheses (some kind of algorithm or a simple function) that when used, can perform the task T.
Now that we know what a learning problem is, let's dive into the types of Machine Learning. Most kinds of learning can be divided into 3 sub-groups: Supervised Learning, Unsupervised Learning, Reinforcement Learning. Supervised Learning uses labeled and structured data sets for training whereas unsupervised learning does not. Labeled data sets have input and output variables defined. Reinforcement learning does not use any training data sets at all.
Supervised Learning
Even within supervised learning, there are two kinds of learning: Inductive and Analytical learning.
Inductive Learning:
- Needs large sets of training data
- Does not require prior background knowledge or rules (often called domain theory)
The following rules can make a domain theory:
Cup <= Stable, Liftable, Open vessel
Stable <= BottomIsFlat
- Basically, statistical inference to fit the training data
- Example: Simple regression, Decision Tree Learning, Neural Networks, Genetic Algorithms
Analytical Learning:
- Does not need large sets of training data
- Required domain theory, i.e. background knowledge of rules for the environment
- Logical deduction to fit the training data and the background knowledge
- Ezample: Example Based Learning
Unsupervised Learning
Unsupervised Learning doesn't use labeled data sets for training the task performing hypothesis. The most famous example for this group is Clustering.
Reinforcement Learning
Reinforcement Learning does not use any training data set. Instead it used the reward-penalty concept in psychology. Given there is an agent in a certain states with certain number of actions available for him to perform, which action should it choose? The agent should choose a sequence of actions that maximize the cumulative reward at the end.
