Machine Learning is a method of data analysis and a branch of artificial intelligence based on the idea that systems can learn from data, identify patterns and make decisions with minimal human intervention. The most common applications they support contain categorization tasks (e.g., document triage), recognition tasks (e.g., image recognition), and prediction tasks. (e.g., engine failure likelihood).
Unsupervised Learning In unsupervised learning, training data does not include labels. The algorithm groups the data into categories or classes of data that are similar in certain ways, which is particularly useful to help understand characteristics of a data set. If a user had a collection of emails and no trained categorizer, an unsupervised approach could group the emails in a way that minimizes the differences between items within a category and maximizes the differences between categories.
In supervised learning, training data includes the desired output, often a category designation, and the supervision comes from comparing the algorithm output with the correct answer. When observed output does not match the desired output, the values from the error condition feed a process that adjusts the values in the algorithm to arrive at the desired output. To continue the email example, the training set would consist of a collection of emails divided into spam and not-spam. The algorithm uses these designators to arrive at the right learned values to give the correct answer.
Reinforcement learning differs from standard supervised learning in that while it has no correct or incorrect answers, it does have a method to calculate the reward for an answer. An example of reinforcement learning is game playing applications. For any move, there is no correct answer; however, the choice of a move will lead to winning or losing the game. The moves that led to the outcome contribute positively or negatively to the reward values for the candidate moves, which the system uses in the next game to improve the outcome.