Machine learning allows computer programs to master new tasks without line-by-line programming. In other words, the programmer is able, to some degree, to allow the software to write itself. AI programs use unsupervised machine learning to make decisions and change behavior based on experience.
How Does Machine Learning Work?
Programmers use machine learning to “train” an artificial intelligence program to perform a function or task, or to learn to reach conclusions based on evidence. The use of this type of training reduces the requirement for extensive programming, which provides step-by-step commands outlining the exact actions to take in each individual circumstance. Machine learning types include supervised, partially supervised, and unsupervised. Unsupervised machine learning gives the AI program the freedom to experiment, determining the most effective methods to achieve the intended result. Unsupervised learning is generally used to classify items into groups or choose appropriate actions based on previous results.
Using Unsupervised Learning for Classification
One form of unsupervised learning involves the sorting of groups of items that fit into a particular category. By comparing data, the AI program is able to find similarities among each data set. This allows the program to sort the entries into groups as needed by the programmers. Gathering data, and sorting it in this fashion, is particularly helpful to business intelligence efforts, which add value to a business through categorizing and analyzing patterns in data.
Teaching an AI Program to Make Decisions
The other form of unsupervised machine learning allows programmers to teach AI programs how to make “good” decisions. The AI program receives reinforcement for correct classification of items during the training phase. The software stores the results of all attempts, including all feedback received. Unsupervised learning methods force the program to base future attempts on past actions, learning from established failures or successes.
Unsupervised Learning Benefits
Unsupervised learning allows the AI program to determine the best, most efficient or most effective means of solving a puzzle, reaching conclusions, or even playing a game, without the limitations of human error or prejudice. TD-gammon, an AI program utilizing neural networks and unsupervised learning, is an excellent example of this type of successful algorithm.
TD-Gammon is an AI program that learned backgammon strategies by playing itself repeatedly. By 1995, it was already earning the praises of multiple world-class backgammon champions, according to Gerald Tesauro, who wrote the program. In his article, “Temporal Difference Learning and TD-Gammon,” which appeared in the March 1995 edition of Communications of the ACM, Mr. Tesauro outlines the various successes of his program.
Training Through Experimentation
The use of experimentation, or trial-and-error, to learn a new task is highly successful for AI programs. By evaluating previous efforts, as well as all data sets given, AI programs can learn new methods of accomplishing these tasks. While the results are not always exactly as expected, unsupervised learning reduces programming time, and can be a viable option for many AI programs.
Cios, K., Kurgan, L., Pedrycz, W., Swiniarski, R. Data Mining: A Knowledge Discovery Approach. Springer, US. (2007).
Information Research Management Association. Machine Learning: Concepts, Methodologies, Tools and Applications. IGI Global. (2011).