Warning...my knowledge of AI, learning systems and heuristics subjects are limited, so please bear with me.... I was thinking about how AI could be built using certain algorithms and/or in conjunction with neural networks or other "learning" based AI systems. I then remembered back to my logics and philosophy class the term heurestics. As I remember it, heuristics in a nutshell is about detecting patterns, and using those patterns to solve problems. That''s when I realized there may be an issue with AI systems that use learning approaches. For example, take the common heurestics question: "are there more words in the english language with ''k'' as the third letter in the word....or are there more words that begin with the letter ''k''?" Turns out the answer is the former, but the majority of people answer the latter because it is easier to think of examples that fit the latter. Our learning and cognitive process use this heurestic as a shortcut to give us an answer. AI can obviously be inflicted with the same limitation as well. As another example, suppose you create an algorithm for AI to use that follows a certain flow-based decision. The AI tests for things when necessary, and depending on the result, will proceed to the next action as appropriate. In a simple non-fuzzy situation, I guess it doesn''t matter, but what if the decision test IS fuzzy, AND you have learning based systems in the mix too? Isn''t it very possible for AI to learn wrong and inappropriate behavior due to bad choices, poor heuristics, or simple chance? In reality, this is quite true, and is how "bad habits" can be picked up. So whether this is a design flaw or not is debateable, but I thought it curious nonetheless.

what you''re describing, in general terms, is biasing the search with context and perception. in implementation this could be as simple as weighting the branches in a tree.

this ties in with learning bad habits is that if an agent hurries down an approach without "thinking it through" properly, whatever learning algorithm is in place will weight this method based on performance which is likely to be POOR, giving feedback.

there is a design flaw if the agent gets the wrong feedback, reinforcing a poor decision, but in a well-designed learning syste this wont happen and the result should be if an agent occasionally makes a balls up of the job, it will try something different next time, which is a natural behaviour.

********


A Problem Worthy of Attack
Proves It''s Worth by Fighting Back