Docstoc

Software Agent __ __

Document Sample
Software Agent __ __ Powered By Docstoc
					Software Agent
인지 구조 4주차 : 제 1 발제 인지구조 / 발제자 : 최봉환



                   John R. Anderson,
          "Human Symbol Manipulation Within
         an Integrated Cognitive Architecture,"
   Cognitive Science, vol. 29, no. 3, pp. 313–341, 2005
Outline

 • Introduction
 • ACT-R
 • Use of brain imaging
 • The capacity for re-representation: A uniquely human trait?
Introduction

 • Overview of ACT-R theory
   – illustrative application of it to algebra equation solving


 • Algebra equation solving
   – uniquely human cognitive activity
   – "what is unique about human cognitive?"


 • Comparing human brain with ACT-R
   – preliminary mapping ACT-R component to brain
      functional fMRI
ACT-R Theory

 • ACT-R
   – Adaptive Control of Thought–Rational
     = cognitive architecture


 • Theory
   – for "how human cognition works"
ACT-R Architecture

 • Role
                                                  Input = Problem
                                                   representation
                                                     (3x - 5 = 7)
      Output
      (x=4)
                                                  massive parallelism &
                                                   central bottle neck
                                                      Mental
                                                  representation
  Retrieve Critical
                                                     (3x = 12)
    Information
     (7+5=12)
                                                  Communication,
                                                 Procedural Control
                      Goal : Strategy decision
                        (unwind stratage)
Algebra equation manipulation

 • Why algebra equation solving problem
   – substantial complexity
   – tractably characterized and studied
       • unlike many human accomplishments
         (cf : Natural language)



 • Problem
   –
                           

   – solved by unwind strategy
The ACT–R model

 • General instruction
   –


       




       
The ACT–R model : speedup

 • Speedup
   – Compilation
      • collapse multiple steps
        into single step




   – Reduction of retrieval times
      • subsymbolic learning
          – instruction strongly encoded during day0
      • arithmetic fact repeated  major learning happening at the symbolic level
          – production rules
Regions of interest

                                          Paretal
                                       problem state
   motor
                                       or imaginal
  manual




  prefrontal                           Caudate
   retrieval   Anterior cingulate
                                      procedural
                      goal
Measuring activity

  • Measuring activity
    – BOLD : blood-oxygen-level-dependent
       • measure neural activity directly have been attempted


    – profile of activity in modules

                                  

       • t = time, s = scales the time,
         a = determines the shape of BOLD response,
         m = govern magnitude
       • f(x) = engage function
Characterizing the differences among
the brain regions
Assessing goodness of fit

  • Measure the degree of mismatch against the noise in the data
    –             
토의 제안

 • 인간과 동일한 구조를 모사하는 것의 의미는?
   – 인간과 동일할 필요가 있는가?
     • 인간에게 원하는 것과 컴퓨터에게 원하는 것이 다를 것 같은데..


   – 인간과 동일한 것을 증명할 필요는 있는가?
     • 1+3 = 4 = 2+2=4라면 내부구조의 의미는?


 • 성능은?
   – 간단한 문제라서 잘 풀리는 것이 아닌지?
   – 수학적인 문제 혹은 논리적인 문제에만 적용 가능한 건 아닌지


 • 모호함에 대한 해결책은?
   – ACT-R은 Deliberative Agent인듯한데 모호한 정의에 대한 묘사는 어떻
     게?
   – Goal based Agent로 구성되어 있는데 목적지는 어떻게 찾을 것인가?

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:2
posted:5/3/2013
language:
pages:13