노홍찬

 ACT-R 5.0: An Integrated Theory of the Mind  J.R. Anderson et al., Psychological Review, 2004, 633 ◦ ACT-R architecture overview ◦ The perceptual-Motor System ◦ The Goal Module ◦ The Declarative Memory Module ◦ Procedural Memory  Reflections of the Environment in memory  J.R. Anderson et al., Psychological Science, 1991, 374 ◦ Form of the memory functions ◦ Environmental Explanation ◦ Formulating the Effects of Practice and Retention

 ACT-R ◦ Adaptive Control of Thought-Rational, first proposed in 1998  Motivation of ACT-R ◦ An image of the mind as a disconnected set of mental specialties. ◦ “how is it all put back together?”  Goals ◦ Producing a theory  that is capable of attacking real-world problems  that is capable of integrating the mass of data from cognitive neuroscience methods like brain imaging

 visual module ◦ for identifying objects in the visual field  manual module ◦ for controlling the hands  goal module ◦ keeping track of current goals and intentions.  production module ◦ coordination in the behavior of other modules ◦ only respond to a limited amount of information in the buffers ◦ recognize patterns in these buffers and make changes to these buffers  serial vs parallel processing ◦ the content of any buffer is limited to a single declarative unit of knowledge, called a chunk ◦ only a single production is selected at each cycle to fire. In this

 Two separate modules for the visual module ◦ Visual-location module with a buffer  Where is the object? ◦ Visual-object module with a buffer  What is the object?  Object identifying process ◦ Request for where system with a series of constraints ◦ Where system returns a chunk representing a location meeting the constraints ◦ Request for what system with the chunk representing the visual location ◦ What system shifts attention to the location ◦ What system generates a declarative memory chunk representing the object  Dual tasks for motor modules ◦ Ex) visual motor, manual motor, … ◦ For the same modules, serial execution ◦ For the different modules, parallel execution

 Responsibility ◦ in the absence of supporting external stimuli keeping track of what intentions are  Keeping a representation of a set of subgoals  Keeping track of problem state

 Cognitive core of ACT-R along with procedural system  Activation of a chunk (Ai) ◦ Base level activation (Bi)  Reflect general the memory’s usefulness in the past ◦ Associative activation (∑W j S ji )  Reflect its relevance to the current context  Wj reflect the attentional weighting of each element (elem j) of current goal  Sji are the strengths of association from each element (elem j) to chunk I ◦ Retrieval probability and the latency is determined by Ai

 Base level activation Bi of chunk i ◦ reflects the log odds an item will reoccur as a function of how it has appeared in the past. ◦ where  frequency of the retrieval of the chunk is n  jth retrieval of the chunk represents jth element of the series  tj represents the elapsed time since the jth retrieval of the chun  many applications suggests the value of the parameter d as 0.5  Bi has been suggested by the author’s previous work  The most successfully and frequently used part of the ACT–R theory.

 The attentional weighting Wj ◦ Wj is 1/n  where n is the number of elements consisting chunk i  The strength of association Sji ◦ Sji is S – ln(fanj)  Where  fanj is the number of chunks associated to element j  S is a parameter, which is estimated as about 2 in many applications  Ex) A hippie was in the park  Each oval is a chunk  Each element has the same attentional weight as 1/3  Hippie in park  Wj is 1 or 3 for each element

 Retrieval probability ◦ Almost the same as Ai ◦ just transformed by sigmoid function ◦ Where  ζ is the threshold that is the minimum Bi for the retrieval to begin  S is a parameter whose role is the transform noise  0.4 for many applications  Latency of the retrieval ◦ just the same as the value of Ai without log function ◦ F is the latency factor

 Activation level calculated by the activation equation  Real retrieval time vs estimated retrieval time ◦ The retrieval time is estimated by the activation level presented above  The correlation between the real and estimated time ◦ ◦ nearly no dependency with the parameters

 The production system ◦ can detect the patterns that appear in these buffers and decide what to do next  Because of the seriality in production rule execution, only one can be selected ◦ the one with the highest utility  where ◦ Pi is an estimate of the probability that if production i is chosen the current goal will be achieved, ◦ G is the value of that current goal, ◦ Ci is an estimate of the cost ◦ Pi and Ci are learned from experience with that production rule.

 Anderson et al, Reflections of the Environment in memory, 1991 ◦ Gave the foundation of base-level learning equation to ACT-R theory  가정 ◦ 인간의 기억 메커니즘은 인간의 진화과정을 통해 환경적인 조건에 최적으로 반응하도록 적응해 왔다.  기존의 연구들은 사람에 대해 적절한 입력을 주고 사람들의 행동을 관찰함으로써 기억의 메커니즘을 찾아내려고 시도함 ◦ 이와는 반대로 인간의 행동이 환경적인 조건에 최적으로 적응하도록 진화해왔다는 가정하에, 환경적인 조건들을 관찰함으로써 인간의 기억 메커니즘을 밝히려는 시도를 수행함  기존의 연구들은 retention function 과 practice function, spacing effect 에 대해서 부분적인 설명이 가능할 뿐, 모두를 다 설명하지 못함 ◦ Anderson 은 환경적인 조건을 관찰하고 이를 인간의 기억 메커니즘을 설계하는데 적용함으로써 위 3 가지 effect 를 모두 성공적으로 설명할 수 있는 theory 를 주장

 Exponential function

 Power function

 Odds ◦ If the probability of event i’s happening is p  the odds is defined as p/(1-p)  Ranging from 0 to infinity  환경에서의 needs odds 가 과거의 기억 인출 기록에 의해서 어떻게 영향을 받는지 알아봄  3 environmental sets ◦ New york times headlines  ~  Checked each word occurrence  Every time a word appears in the text, it’s a request for the reader to retrieve the word ◦ CHILDES database  Related to children’s verbal interactions  Every time someone says a word to a child, it’s a request on the child to retrieve the word’s meaning ◦ Mail messages of the author  ~  Every time the author receives a message a certain person, it’s a request for the author to retrieve the memory of the sender

 Recency effect

 Frequency effect

 Spacing effect

 Basic assumptions from the environmental experiments ◦ The strengths from individual presentations sum to produce a total strength (frequency effect) ◦ Strengths of individual presentations decay as a power function of the time (recency effect) ◦ The exponent of the power function for decay of each presentation decreases as a function of time since previous presentation (spacing effect)

 Base-level activation vs Cache algorithms ◦ Base-level activation  Frequency effect 와 recency effect 의 결합  Power function 의 활용 ◦ Cache algorithms  Need odds vs cache replacement policy  LRU, LFU, LRFU (1999)  LRU: only recency effect 고려  LFU: only frequency effect 고려  LRFU: recency effect + frequency effect  not with power function

 Chunking using ACT-R vs association rule mining ◦ Association rule mining  Offline algorithm ◦ Chunking with ACT-R  Online algorithm