8 Jun 2009 mentifex   » (Master)

Cognitive Chain Reaction for AI Troubleshooting

#summary Ideas looping in an endless cycle

MileStones RoadMap UserManual

=== Definition === 

A cognitive chain reaction (CCR) in a nascent AGI can be defined as a series of three or more natural-language thoughts which, in the absence of cognitive distractions, enter into an apparently infinite loop. By the process of SpreadingActivation, each thought in the loop leads to the next thought, which in turn leads to the next, and to the next, and so on, _ad infinitum_.

=== Attainment ===

Long after the nascent AGI has achieved QuIckening, and shortly after the AI Mind has achieved GenerationOfThought, an AGI coder will induce a cognitive chain reaction by entering a series of looping ideas such as the following.

{{{ Cats eat fish. Fish eat bugs. Bugs eat germs. Germs kill cats. }}}

Upon entry of the last of the looping thoughts, the AGI will associate from the last word entered back to the start of the loop, and will begin an endless repetition of the loop.

=== Purpose ===

The cognitive chain reaction serves mainly as a troubleshooting device, to make sure that after changes to the AGI source code, the AI Mind is still able to think. Rather than labor to dream up items of novel input to test the AGI, the programmer enters a tried-and-true series of looping thoughts for a quick assurance that the associative mechanisms still work properly.

The actually quite mundane purpose of the cognitive chain reaction is not to achieve any sort of gee-whiz "wow!" effect, but simply to verify that the AGI can think in a repeatable set of circumstances.

=== Upshot ===

In practice with the MindForth AGI, while being a valuable diagnostic tool, the CCR has at times generated weird but still acceptable results. Until the final thought in the loop is entered, the AGI typically greets each new idea with a question, such as "BUGS...WHAT ARE BUGS?" Such a response is actually quite sophisticated, because the AGI starts to generate a thought about "BUGS" but does not know any verb for completion of the thought. A module kicks in to ask a question about the new concept about which the AGI has no data in its KnowledgeBase.

Because the chain of thought spreads or loops only if there is sufficient activation to keep it going, sometimes there is a kind of conceptual "hiccup" in the infinite loop, where the AGI pauses to ask a question about one of the words and then answers its own question by continuing the loop. To the AGI programmer, such behavior is an indication that the pre-programmed parameters of conceptual activation may need some tweaking or refining. Typically, the coder is so eager to get on with other tasks that the tweaking is put off for a later date.

Another weird result occurs when the AGI cannot recognize a word like "fish" as being in the plural. The AGI will tend to say "FISH EATS BUGS" as part of the loop, instead of "FISH EAT BUGS." But the loop continues.

If you are demonstrating MindForth or a similar AGI in ScienceMuseums or at a robotics club, it is instructive to show that that the cognitive loop may be interrupted at any time by novel input from a human user.

=== OutReach === http://agi-roadmap.org/Cognitive_Chain_Reaction is an open, collaborative page where you may contribute your own ideas and use the associated Talk page for discussion.

=== MemeSpace ===

AdminisTrivia AiEvolution AiHasBeenSolved BrainTheory CognitiveArchitecture ComPutationalization ConSciousness DiaSpora DisAmbiguation EmBodiment EnTelEchy ForthMindTextFile GenerationOfThought HardTakeoff HumanLevel ImMortality IndustrialEspionage InFerence JavaScript JointStewardship KbSearch KbTraversal KnowledgeBase MachineLearning MasPar MeanderingChain MetEmPsychosis MileStones MindModule MovingWave NaturalLanguageProcessing OutReach PermanentInstallation PortingOfCode ProliFeration ProsperityEngine QuIckening RecursiveSelfEnhancement ReJuvenate RoadMap RoBot RumorMill ScienceMuseums SeedAi SelfReferentialThought SemanticMemory SeTi SpreadingActivation SubConscious SuperComputer SuperIntelligence TechnologicalSingularity TelePresence TimeLine UserManual VpAi WikiPedia

labels: debug milestone roadmap

Latest blog entries     Older blog entries

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!