This document is available in a printable pdf form: SPLASH: Goal Oriented Machine Learning Handout
Goal Oriented Machine Learning
Machine learning is a subfield of the field of Artificial Intelligence, which is a subfield of both Computer Science and Cognitive Science.
Parts of a Machine Learning Computer Program
So what are different parts of computer programs that can be replaced by learning algorithms? Well, since there are many different types of computer programs, let's try to make a list of the different types of parts of each of a sample list of different types of computer programs that currently exist in the world.
| Example Application of Learning | Learning Algorithm Used |
| Music Recommendation System | Eigenvector and other Linear Techniques |
| Spam Filtering | Probabilistic Bayesian Network Classifiers |
| 4-Wheel Drive Antilock Braking System | Feedforward Backpropagation Neural Networks |
| Polyworld: Artificial Life Evolution | Genetic (and other Search) Algorithms |
| Simple Animal Psychology (Reward/Punishment) | Reinforcement Learning Algorithms |
| Credit Assignment in Plan Failure | Causal Reflection Algorithms |
| OpenMind Commonsense | Being Told |
| Emotion Machine One (EM1) | Analogy |
What is a Goal?
Every part of a computer program that can be learned automatically has two parts: (1) static, and (2) dynamic. The part of the program that is programmed by the programmer and is static is called the learning algorithm or process, whereas the part of the program that is dynamic is the part that adapts to the run-time experience of the program is the representation. A goal is a state in your representation space that you would like your computer program to obtain. For example, if there is a part in your computer program that you would like to be handled automatically, then that part can be learned by a learning algorithm. How does a learning algorithm get closer to the goal performance? A learning algorithm has experiences that it learns from. The more experience that a learning algorithm has, the better or closer that algorithm comes to accomplishing the goal. There are many different types of goals and there are many different types of learning algorithms that handle each of these different types of goals! Basically, for every different part of a computer program that a person can write, there is a different type of learning algorithm that can replace that human designer with a learning algorithm that learns from enough experience.
| Static (Programmed) Algorithm | Dynamic (Learned) Representation | Goal Representation |
| Eigenvector and other Linear Techniques | Normalization (Whitening) Matrices | None (Information Compression) |
| Probabilistic Bayesian Network Classifiers | Conditional Probability Distributions | None (Predictive Accuracy) |
| Feedforward Backpropagation Neural Networks | Numerical Weights for Each Synapse | Supervised Training Samples |
| Genetic (and other Search) Algorithms | Recombinant Data (Strings) | Fitness Function (and Heuristics) |
| Reinforcement Learning Algorithms | Reward Model and World Model | Active Reward and Punishment |
| Causal Reflection Algorithms | World Model and Planning Model | Declarative Structures and Failure Types |
| Being Told | None | None |
| Analogy | None | None |
| Experimentation | None | None |
When should we use Machine Learning?
Why wouldn't we use a learning algorithm in order to replace all parts of a computer program so that people don't have to design any of the parts? Well, each part of the computer program that is a learning algorithm, must have enough experience to learn from so that it can learn to function as well as the same part of the program that could have been designed by a human programmer. In other words, for every part of a computer program that is a learning algorithm, there must be a lot of experience in order to train that part of the program. Also, sometimes learning from experience is not a good idea because the goal that is to be accomplished is very important or dangerous. For example, if the algorithm is learning to control a nuclear power plant cooling system, a learning algorithm that learns under what conditions the power plant melts-down would not be a good part of the program to be learned from experience. Instead, important or dangerous parts of computer programs should not be learned by a learning algorithm that must learn from experience.
