0203. Introduction to Artificial Intelligence
Issues Revisited
In the following, many issues discussed in the course are revisited, mainly according to the solutions explored in the NARS project. We will see that it is possible to solve these issues in a unified manner.
Here the key is not to demand a predetermined algorithm for all instances of a problem type, but to handle them case by case, under the restriction of available knowledge and resources.
- [All options are listed at each step] Instead, let the inference process constantly establish new possibilities.
- [There is a predetermined heuristic function] Instead, the comparisons are based on revisible truth values.
- [In each step, a choice is made among several options] Instead, the processing time is unevenly distributed among options.
- [Each operation leads to a certain state] Instead, the consequence of each operation is only certain to a degree.
- [The process stops at a goal state] Instead, the process usually stops when it has no resource.
For the uncertainty in the meaning of a term, the key is to see it as determined by the experienced relations with other terms.
- [The meaning of an atomic term does not change as the system is running] Instead, new and derived beliefs constantly change the meaning of the terms involved.
- [The meaning of a compound term can be reduced to its components] Instead, the relations between a compound and its components only consist of part of its meaning.
For the uncertainty in the truth value of a statement, the key is to see it as a function of available (positive and negative) evidence.
- [A statement is either true or false] Instead, truth value is a matter of degree.
- [The truth value of a statement does not change over time] Instead, new evidence always changes related truth values.
- [All beliefs must be consistent] Instead, beliefs may have explicit or implicit contradictions among them. In a term logic, a contradiction does not imply an arbitrary conclusion.
For the uncertainty in the process of inference, the situation is the same as discussed in search. As far as the process does not follow a predetermined algorithm, it will be as uncertain as the human reasoning process.
All the inference rules, deductive or not, can be justified according to the experience-grounded semantics, that is, the truth value of a conclusion indicates the evidential support it gets from the premises, not its agreement with an "objective fact".
Most of the traditional "paradoxes of logic" are actually "paradoxes of first-order predicate logic", and they disappear in a properly designed logic.
- [Sorites paradox] Since "heap" is a fuzzy notion, each grain of wheat changes the degree of membership a little bit, and the effect accumulates.
- [Implication paradox] When "If P, then Q" is taken to be a variant of the inheritance relation, its evidence must link P and Q in contents.
- [Confirmation paradox] "Ravens are black" and "Non-black things are not Ravens" have the same negative evidence, but different positive evidence. Therefore, they are equivalent only in a binary logic.
- [Wason's selection task] If truth value is determined by both positive and negative evidence, it is valid to look for confirming cases.
3. Learning
All empirical knowledge, such as the meaning of a term, the truth value of a statement, and the usefulness of a concept, are all determined by the system's experience, rather than by the designer.
Learning does not follow a predetermined algorithm, but is a life-long process influenced by many events in the system.
The learning process and the reasoning process of a system can be unified. In this process, the system uses justifiable rules to produce new terms, new statements, and new truth values.
Learning is a statistical process, in the sense that the meaning of a term, the truth value of a statement, and the usefulness of a concept are all determined by many events, rather than by a deterministic step.
Learning is restricted and regulated by resources competition. In the long run, only useful concepts and beliefs will survive.
4. Knowledge representation
AI needs a knowledge representation that is general, expressive, modular, natural, and flexible. It should uniformly cover knowledge of different types and domains.
The knowledge representation should support efficient knowledge processing.
The knowledge base of an AI system should not require detailed human management.