Toolkit

Story: Artificial Intelligence: Working backwards from HAL

• 
• 

Reasoning Systems

In my opinion, the future of AI is dependent on the extension of the programming paradigm from today’s Boolean logic to incorporate “reasoning models”. Compsim’s KEEL® Technology offers one approach. Only when systems have the opportunity to exercise “reason” will they evolve to the next level (HAL-like). Reasoning is not a sequential process. It is an analog balancing process where inter-related alternatives need to be considered. It requires the production of relative answers and actions in dynamic environments.

I believe there are two competing mindsets driving the future:

First there are researchers with the “mechanism” mindset that use biological models as the foundation for their work in AI. They focus on neural nets and genetic algorithms that will allow applications to learn on their own and evolve on their own. They attempt to model how the human brain functions. Most of the research seems today seems to be focusing on this approach. The risk with these systems is that they may evolve in directions never before expected, even by their designers. If they create human-like reasoning engines, they can evolve in good ways and in bad ways. Just look at humans.

The other mindset assumes that there must be another way to create “reasoning systems”, with the added demand that they must be completely explainable and auditable. Humans must retain control. With this approach one is searching for a solution, but is not tied to the biological model. I call this the “process” mindset.

Without worrying about the technical details about how a human “reasons” or makes “judgmental decisions”, the reasoning process is commonly understood to take place in the human’s right brain. The left brain may focus on language and logic (the domain of most computers today). The right brain performs image processing and makes judgmental interpretation of those images. The images are not necessarily “pictures”, but can be feelings or impressions.

This suggests that text based programming languages do not satisfactorily provide a platform for “reasoning”. Similarly, I would suggest that it is difficult or impossible for humans to explain “exactly” how or why they make judgmental (subjective) decisions. Just watch the news broadcasters attempt to explain why humans do what they do. Or watch CEOs explain why they did what they did.

“Humans”, each perform their own “interpretation” of information. They provide their own weights to supporting and objecting arguments. They fuse the differing viewpoints in different ways. Human language does not allow this “reasoning model” to be effectively exchanged. It is for this reason that (industrial) machines are not commonly controlled with human language terms. They are controlled with numerical settings and formulas that tell machines exactly what to do.

Fuzzy logic is one design model that attempts to bridge human linguistic terms to the machine model. It starts to bring graphical constructs into the solution. Complex fuzzy logic systems, however, may be difficult to design and diagnose.

Creating solutions that can exercise reason is still the objective. Whether the market is for personal robots to take care of the aged, for automated medical treatment to reduce the cost of health care, or for robotic weapons to fight future battles, these systems must be able to exercise reason to be effective. New languages (like the KEEL dynamic graphical language) will be required to define the reasoning models. New engines (like those based on KEEL designs) will be required to process the reasoning models.

There is one thought that the future of AI requires larger and faster computers. I would suggest this is “nice”, but not necessarily required for “reasoning”. We don’t necessarily need millions of HALs running around (at least for the near term). The industry will evolve by building focused systems that exercise reason and address focused

• 
• 

• 
• 
• 

Full Talkback thread