This guide is on how to train and use Machine Understanding.
The first part speculates search as the main function of intelligence. The second offers training principles that make the use of such intelligence effective. The third discusses an example of a training program that utilizes these principles.
Part One. Learning Search Agents
The main function of an intelligent agent is the search for a solution within the boundaries defined not only throughout the training process but when already in use. This difference to traditional AI is important, as what must be trained is not a utility function but an ability to build utility functions of some kind. Part One demonstrates the idea of making intelligent search agents through applying machine learning methods to algorithms of random search.
Part Two. The principles of Machine Understanding
To ensure that the agent’s learning is effective, Part Two offers four basic principles of training. These principles are double-sided as any training uses some properties of the learning agent, and has to comply with some properties of the subject domain. Despite the loss of universality, the proposed training process applies practically in any subject domain developed by humans, since human thinking seems to be based on the like principles, and the very knowledge is possible only due to these properties of the subject domains.
Part Three. Training to Understand Arithmetic
Part Three describes the general structure of training programs for Machine Understanding, which is demonstrated on the example of training numbers in Arabic notation and arithmetic operations.
Appendix. Notes on Training AI to Program
Making an algorithm is a universal way to describe a problem solution. Training to create algorithms is the ultimate goal for AI as it differentiates mere memorizing from understanding. The notes are on why this kind of training is possible despite some interpretations of some math theorems that seem to prohibit searching for algorithms.