Share Site Map Contact us Home Page




Alan

HAL
Video News Release
The Hal Nursery
HAL 2003
HAL Personalities
Terms of Use

MyBot
HAL 2003

HAL consists of two basic components: The Brain and the Personality. The next generation of Brain, V20.0.3 is yet another step forward in the evolution of language acquisition machine learning algorithms developed at Ai.

HAL 2003 differs from its predecessors in one major aspect: All previous versions were flat: single context algorithms. The Brain consisted of a set of single-variable rules, competing for the right to issue the current response. This mechanism included no notion of CLASSES: all rules were assumed to pertain to any value of the variable. For example: The rule

U> What is a [VAR]?
H> A [VAR] is a color.

Would produce this behavior:

U> What is a dog?
H> A dog is a color.

In HAL 2003, each rule is associated with a CLASS, of all the permitted values of [VAR]. Thus, if the class associated with the rule in the above example is {red, blue, green, yellow}, then 'dog' would not be a member of this class.

The one-to-one correspondence between rules and classes, introduces a hierarchy in the rule-system, as opposed to the previous, flat, non-hierarchic scheme. This new feature of HALs brain constitutes another step forward in its learning capability, enabling the automatic propagation of rules from sub-class to super-class.

The new version uses the same heuristic engine but with a different (yet similar) set of heuristics. The following changes should be noticed on the new engine:

1. Smaller personality sizes -> yielding faster load and save times.

2. Smaller number of heuristics -> yielding faster response times.

3. The Magnetico (info-magnetism) heuristic has been removed -> the 'signature phrase' of a personality will change and hopefully will go away all together over time.

4. The usage of a 'one turn deep' deeper context has been removed. This is been replaced by a mechanism which associates events in the past with the current context in a way the enables the creation of 'state variables'. This is a major change and more on that later.

5. The flow of learned material between classes of sub/super relation has been implemented. This means that once a class is established, new material needs not be taught for all of its members.

In general, the engine adds the usage of the class members and relations to the logic associated with rule generation. The classes can be used to share behavior (as is the case of a sub/super relation) and as a sort of a state variable (using the last used element of a class).


HAL 2003 is currently deployed alongside the previous version. All new personalities (based on a 'Fresh Copy') utilize the new engine, while old (private and public) personalities utilize the previous version.