Truth, Justice, and the American Way

Proposal for an Evolution-Based Adaptive Artificial Intelligence Algorithm for Conceptualizing Natural Language Terms using a Competition/Cooperation Multithreaded Population Simulation Model

v.2

The basic idea is to simulate a herd of simple brains and create an environment that makes them evolve into intelligent minds

The "food" is successful creation of algorithms that automatically solve a given problem set and improve and create new algorithms for processing ever more complex problems.

This evolution occurs by random process mutation, “mating” between two separate AI lives as well as information sharing between them.

The AI life forms will learn to cooperate and solve problems together by sharing information and distributing parts of problems.

Eventually concepts would be in plain English and they could process them just as a human mind does using a conceptual model that uses simple logical statements to evaluate and process language.

Two primary components:

Genetically and Socially Adapting Evolutionary Mechanism:

The AI life forms will interact with each other by A. Mating with each other to create offspring, which combine the parent’s genetic and social traits as well as a random factor, and B. Sharing and eventually trading information.

A combination of genetic and social traits that use fuzzy logic to create ranges for each social and genetic characteristic will set traits such as times between mating, and sharing information, and the distribution and use of “digested” food obtained from processing problems.

The key and most difficult part of this part is to create the ability to learn complex new behaviors that facilitate information trading and cooperation rather than simple sharing of collected data.

Concept-creation engine:

This is the more difficult part of the program. Each AI life form will have an independent “mind” that will use inherited and learned rules for processing data. They will use their mind to solve problems given to them by the programmer and gain “food” by successfully solving them or lose stored food by failure. This will go on until the life form either A. Has enough food to reproduce, (and start over) or B. Runs out of food and dies.

Depending on its genetic traits, it will at certain periods attempted to share information with other life forms.

For example, if the challenge is “Are cats mammals?” the life form will access the records “Mammals: subset(dogs, cats, bats) and use the logical question Is (C) = (M)? Since C is a subset of M, the statement is true, and correct reply is “Yes” The life form will then be rewarded with food, and soon on.

If the question is “What do you think of President Clinton?” initially the life form would have to guess between “Good and Bad” (or other possible answers to a “what do you think” question) Since “Bad” is the proper answer, a Clinton(Bad, President, Man, Name) entry could eventually develop. Perhaps probabilities could be added to the entries such that Clinton(Bad:70%, President:40%, Man:99%, Name:95%) in case it was sharing information with another AI and saw that it has a Clinton(Good) entry. (Programmer error in this case)

Eventually, there would develop a large massively relational database relating all the terms. The challenge then would be to develop concept-models for processing new and undefined concepts. For example if the question is “What year was it 10 years ago?” the lifeform might try guessing at first, first random answers, then after learning that Year(Number) random years. Eventually it would try a numerical analysis in the form of Year =X-ThisYear IF (was) and (ago) Then it would have the idea of time as a concept. This is the hardest part, as it would have to have the ability to come up with these concepts all by itself.

Eventually, there would be concepts for a wide enough vocabulary that the program could be given a text and then be able analyze that text and come up with the relevant concepts, maybe interactively asking a human to explain the ones that contradict or are not complete.

Since the world (and English language) is not a neat little ball making perfect sense, the AI would need alone time to process, organize, and categorize it’s date, otherwise known as an inner dialogue in humans. At this point, it could be said to be intelligent.

Challenges:

How to go from a meta-dictionary to the creation of a concepts.

How to encourage teams and cooperation between lifeforms.

How to create a massively linked database for a whole herd of lifeforms without requiring a supercomputer.

How to have a herd of independent and co-functioning life forms without requiring parallel computing.

Possible References:

Jake Beal: Bootstrapping Communications from Shared Experience

http://www.ai.mit.edu/research/abstracts/abstracts2001/machine-learning/01beal1.pdf

AI Technical Report January 2002: Generating Communications Systems Through Shared Context:

ftp://publications.ai.mit.edu/ai-publications/2002/AITR-2002-002.pdf

A Parallel Computing Approach to Creating Engineering Concept Spaces

for Semantic Retrieval: The Illinois Digital Library Initiative Project