AI Redefined

download AI Redefined

of 37

  • date post

  • Category


  • view

  • download


Embed Size (px)


AI Redefined. Given what we have studied in this course, the author offers a new definition AI is the study of the mechanisms underlying intelligent behavior through the construction and evaluation of artifacts designed to enact those mechanisms - PowerPoint PPT Presentation

Transcript of AI Redefined

  • AI RedefinedGiven what we have studied in this course, the author offers a new definitionAI is the study of the mechanisms underlying intelligent behavior through the construction and evaluation of artifacts designed to enact those mechanismsThere are several noteworthy things about this definitionwe only commit to intelligent behaviorevaluation is a critical component for the definition we are no longer relying just on the Turing Test, but instead we look to evaluate the performance of any AI system although the definition does not tell us how to perform that evaluationthe emphasis here is on artifacts working systems

  • AI as a FieldFrom the definition, we see that AI is (or should be) less concerned about a central or single theory of mindit is more empirically targeted working systemsThus, AI is an engineering pursuit the creation of working systemsThe need for evaluation makes AI a sciencewhile creating systems is well and good, without analyzing the systems to understand why they work or do not work, AI will be working in a voidThis definition denotes a paradigm shift away from philosophy of mind we do not need to study mind to create mind, human mind might help with models but the formal studies found in philosophy have done little to helppsychology of the human mind we may gain some understanding of what to do through experimentation but again, this sort of pursuit has led AI astray

  • PSS Hypothesis ReduxRecall in our first lecture, we considered the PSS Hypothesis a PSS has the necessary and sufficient means to exhibit intelligent actionthe focus on this hypothesis led AI to researchuse of symbols to model the worlddesign of search strategies to apply operators on the given symbolsuse of heuristics to guide the searchthe use of an empirical approach to research build (construct) and test to prove your pointWhile this has helped form a basis for AI, it has also misled AI much as earlier reliance on philosophy and psychologydo we need symbolic knowledge? neural networks show otherwisedo we need heuristic search strategies? model based approaches (whether structural/functional, Bayesian or HMM based seem to indicate less need for this)

  • Why Has AI Not Succeeded?The author continues by examining the challenges ofsymbolic AI lack of grounding of symbols, lack of a social context by which symbols are learned in humans, systems constructed using symbolic AI approaches remain too brittle (striving for single interpretations rather than multiple interpretations or contexts, having limited amounts of knowledge)subsymbolic AI neural network nodes are not equivalent to neurons, the number of nodes differs substantially from a brain, even of the smallest creatures, while neural networks can be used to construct context-sensitive memories, the actual storage of memories (situations, cases, events) remains beyond our abilities because we do not know how memories are formed in the brain

  • What Should AI Research?This is an open questionwhat we tend to find is that there is less emphasis in AI research on a grand unifying theme or holy grail and far more emphasis on solving fundable problemsSome of the more common approaches found in AI research are toexpand the capabilities of neural networks by identifying new algorithms, combining NNs with approaches like GAs and FLagent-based approachesmathematical model-based approaches (for instance, HMMs)ontologies to provide problem independent knowledge sourceslearning from the ground up

  • My Research: AbductionGiven data to explain, search for possible explanaiers (hypotheses)Score themAssemble them into a composite explanation

  • Hypothesis Assembly AlgorithmLook for essential hypothesesa hypothesis that is the only way to explain some dataInclude/propagate/remove (see the next slide) and repeat from topLook for superior hypothesesa hypothesis that is clearly superior at explaining some data because its plausibility value is substantially higher than any other explainerInclude/propagate/remove and repeat from topLook for better hypothesesany hypothesis that explains remaining data better than any otherInclude/propagate/remove and repeat from topIf there are still data to explain, either guess or quit with unexplained data

  • Include/Propagate/RemoveIf a hypothesis is found as essential, superior or better (or guessed at) theninclude it in the compositepropagate the results of including this hypothesisif this hypothesis is incompatible with other potential explainers, remove those other hypotheses this could create new essentialsif this hypothesis is either associated with or lends support to another hypothesis, increase that hypothesis plausibility this could create new superior or better hypothesesIf this hypothesis diminishes support from another hypothesis, reduce that hypothesis plausibility in doing so, this could allow another hypothesis to become superior or betterremove all data that the newly included hypothesis explains

  • Two Concepts

  • ExampleImagine that we have 6 hypotheses available to explain 4 data as shown:

    the solid lines indicate what each hypothesis can explain, the dashed lines indicate hypotheses that are mutually exclusive and the dotted line with the (+) indicates support (if H2 is true, H3 is supported and vice versa)Our best explanation will be generated as follows:select H4 (confirmed), select H1 (essential) which rules out H6removing H6 makes H2 an essential (only way to explain D2)H2 supports H3 so H2 is a better choice to explain D3Best explanation is {H1, H2, H3, H4}

  • Abduction AppliedRed blood cell typing (data interpretation)given blood cell reactions, explain them in terms of blood antigensMedical diagnosis (liver disorder diagnosis)Speech recognitionARTREC input was microbeam pellet data, so this was more like a lip reading systemNatural language understandingTheory formationwhich theory better supports life on Earth, evolution or creationism?which theory better supports evidence produced in a trial, a person was the murderer or not?Hand-written character recognition

  • ARTREC: the data

  • ARTREC: Gestures

  • ARTREC: Noise Hypotheses

  • ARTREC: Abductive Inference 1

  • ARTREC: Word Hypotheses

  • Layered Abduction

  • Evolution vs Creationism

  • Continued

  • Hand-written Character Recognition

  • Explaining a CharacterThe features (data) found to be explained for this character are three horizontal lines and two curvesWhile both the E and F characters were highly rated, E can explain all of the features while F cannot, so E is the better explanation

  • Top-down GuidanceOne benefit of this approach is that, by using domain dependent knowledgethe abductive assembler can increase or decrease individual character hypothesis beliefs based on partially formed explanations for instance, in the postal mail domain, if the assembler detects that it is working on the zip code (because it already found the city and state on one line), then it can rule out any letters that it thinks it foundsince we know we are looking at Saint James, NY, the following five characters must be numbers, so I (for one of the 1s, B for the 8, and O for the 0 can all be ruled out (or at least scored less highly)

  • Full Example in a Natural Language Domain

  • Decision MakingBased on the abduction algorithmused to solve 3 different problemsgrocery shopping list generationmeal planningdepartmental course scheduling

  • Other Research AreasClassificationautomated syntax error debuggingLinux user classificationgrammatical errors of non-native English speakersstudent error classification CAI toolMusic creationcombination of routine design and genetic algorithmsAutomated software creationcombination of case-based reasoning and routine designselect code components from a code library based on function (goals)use pseudocode plans as prior cases

  • Research Topic: The Semantic WebHow can we automate the process of using the knowledge available on the Internet?we need to make the knowledge available in accessible forms (ontologies)we need to provide a suite of problem solvers that canfind the knowledge they needmake decisions based on the knowledge foundcommunicate with other problem solvers when the needed knowledge is not available, or when they have a specific subproblem to spawncommunication might require social interactions beyond simple message passing (for instance, polling of many sources, determining if an agent is trustworthy)migrate to other processors either because their current processor is busy, or more usefully, because the knowledge needed is located elsewhere (this capability is optional)

  • Research Topic: Autonomous VehiclesHow can a vehicle be programmed to carry out a mission on its own with little or no human intervention?requires mission planning, path planning, sensing (and sensor interpretation) decision making, reactive planning, failure handlingall of these steps must be done in real-time except mission planning and path planning which can be done prior to the start of the missionEach type of vehicle has its own unique challengesairplanes and submarines deal with 3-D, have fewer obstacles to contend with, but also have draft/currentautomobiles have to deal with other road traffic, off-road vehicles have to deal with rough terrainindoor robots deal with human traffic, furniture, walls, etcmany of these vehicles do not use cameras for input but sonar and/or radar instead

  • Research Topic: Evolving IntelligenceRodney Brooks from MIT claims that AI needs to evolve through self-learningin his lab, he has a number of robots placed into an environmentthe robots start with a base behavior that is la