Intelligent Systems And Their Societies

LIVE

by Wei-Min Shen, MCC, Austin, Texas

LIVE, just like all ISs, has innate elementary responses and senses. Elementary responses can be such things as a muscle contraction in a biological system or the switching of a motor in a robot. These are parts of the system that generally were fixed when the system was built. LIVE grounds the meaning of its symbols (of all its concepts and response rules) in the systems elementary sense information and elementary responses. If an IS does not do this, the meaning of its symbols cannot be within the system, but only in the human mind that created it. This is a critical difference, because such a system cannot think for itself: it is only able to "manipulate words", do "number crunching", or do other sequences of actions for it's operator.

Learning Engine
Response rules, the linkage of responses to their consequences, are always unknown at the point of entering a new environment. They are then learned from the environment as one of the IS's first tasks. During this early stage of learning, appropriate response rule exploration is very important. In addition, a response by the IS may end up causing a substantial change in an environment. In order to better deal with this, LIVE was designed to be able to learn that the same response can have different consequences in different environments.

Response rules, which Shen calls "prediction rules", have as parts the stimulus (called "condition"), the action, and the resulting situation (the prediction). LIVE expresses the stimulus and the resulting situation (the prediction) as concepts (called "precepts") that are connected by logical connectives such as "and", "or", "not". LIVE's author calls these rules "prediction rules" since they help the system learn to predict the consequences of responses.

As a further benefit of the particular form of LIVE's rules, namely that they include the resulting situation, the system becomes more adaptable to minor changes in its environment. With the sum of its response rules it creates a model of the current environment; the response rules represent the environment. If, however, a prediction based upon this model is suddenly found to no longer be correct, the system experiments with responses until it learns a correct prediction.

As with most things, this same rule form also has a downfall. While an experimenter may manually enter response rules, (for example, to test different situation comparison engines) he may do this only by expressing them within the language of actions and concepts of the system. This is often awkward and sometimes even impossible because of the unknown nature of the interrelations between concepts and rules.

Second Order Learning
The output of LIVE's sense organs are elementary concepts (called "percepts"). No system that does not have an infinite number of sense organs can know all of the important features of its environment. For instance a color blind system cannot know the color of objects. LIVE, however, can sometimes "discover" these "hidden" features when a response does not have the expected consequences.

A goal generator can accept objectives from a person or generate its own. To reach towards an objective, LIVE's goal generator uses a planner to generate its plans. These plans consist of response rules. If, during execution of a plan, a hidden concept surprises the system and causes it to reach an unforeseen situation, LIVE revises the inconsistent rule, sometimes creating new (inferred) concepts.

Results
LIVE was tried in many applications including games (tower of Hanoi) and child learning (creating and learning the concept of "torque"). In the area of scientific discovery it found Mendel's gene laws. These results showed LIVE to be an interesting and improved IS.

For continuos reading, like a book - continue here
Jump to the top of the LIVE document /Artificial IS document /e-book Contents.