INTRODUCTION TO ACT-R

Transcription

1 Cognitive Modeling WS 2012/13 2 Outline Overview and characteristics INTRODUCTION TO ACT-R Eduardo R. Semensati Components Chunks; Productions; Buffers. Modules The Perceptual-Motor System; The Imaginal/Goal Module; The Declarative Memory Module; Procedural Memory. Cognitive Modeling WS 2012/13 3 Cognitive Modeling WS 2012/13 4 ACT-R - Overview Adaptive control of thought-rational The vanilla ACT-R This presentation: The cognitive architecture The theory behind it ACT-R - Characteristics Modularity Each module has a special purpose Two different types of knowledge Declarative (chunks) Procedural (productions) Mixture of parallel and serial processing Within each module: Parallel The two serial bottlenecks Cognitive Modeling WS 2012/13 5 Cognitive Modeling WS 2012/13 6 1

2 Cognitive Modeling WS 2012/13 7 Cognitive Modeling WS 2012/13 8 Components - Chunks Elements of declarative knowledge (knowledge we are aware of) Examples: George Washington was the first president of the United States 5+5 = 10 Components - Chunks Defined by its chunk-type and its slots. Chunk-type: represents a category. Example: birds Slots: represent category attributes. Example: color or size The chunks are displayed as a name, and then slot and value pairs. Name -> Fact3+4 Type -> isa addition-fact addend1 three addend2 four sum seven Cognitive Modeling WS 2012/13 9 Cognitive Modeling WS 2012/13 10 Components - Production Elements of procedural knowledge (generally not consciously aware of) Examples: Speaking a language. Riding a bike or driving a car. Components - Production Statement of a particular contingency that controls behavior. Represented as if-then rules. The Condition (LHS) Patterns of chunks that must be present in the buffers The Action (RHS) Actions to be taken when the production fires. Cognitive Modeling WS 2012/13 11 Cognitive Modeling WS 2012/13 12 Components - Buffers Interface between the procedural memory system in ACT- R and the modules. Modules - The Perceptual-Motor System Nature of cognition is strongly determined by its perceptual and motor processes. The actions of a production affect its contents. Can hold one chunk at a time (or none). This system is formed by two components: The motor module The vision system Visual-location module and buffer; Visual-object module and buffer. 2

3 Cognitive Modeling WS 2012/13 13 Cognitive Modeling WS 2012/13 14 Modules Motor Module Provides the architecture with hands. Modules Vision Module More visual attention than a theory of perception. The approach: modelling the basic time behavior of those motor systems. Kind of actions: using the mouse or typing in a keyboard. It has two parts: Visual-location: the where system. Works with constraints and supports visual pop-out effects. Visual-object: the what system. Shifts visual attention and processes the object there. Two level of granularity (regarding time). Cognitive Modeling WS 2012/13 15 Cognitive Modeling WS 2012/13 16 Modules The Imaginal/Goal Module Previously: only the Goal module existed (ACT-R 5.0). Nowadays: division between Imaginal and Goal module. The goal module: Responsible for keep tracking of the current and next steps/subtask (Control Flow). The Control state. Modules The Imaginal/Goal Module Example: adding two numbers. Add numbers in ones' place; Save result and identify carry, if exists; Add numbers in the tens' place; Save result and identify carry, if exists; Report answer. The imaginal module: Responsible for the intermediate results. The problem state. Cognitive Modeling WS 2012/13 17 Cognitive Modeling WS 2012/13 18 This is the memory responsible for facts. All the knowledge are stored as chunks. Acces to information is hardly instantaneous or unproblematic. Some questions arise: Can a fact be retrieved at all? If yes, how long does it take? In case of conflict, which fact is retrieved? What controls the access? The activation processes! Base level activation Contextual influence 3

4 Cognitive Modeling WS 2012/13 19 Cognitive Modeling WS 2012/13 20 The equation presented represents the activation of a chunk i, and it determines: If a chunk can be retrieved: Which chunk will be retrieved: highest A i And the retrieval time: A i >! RT = Fe! A retrieval threshold latency factor The base-level activation B i Reflects the odds that you need a chunk based on frequency and recency. What is a presentation? Creation of a chunk Buffer clearing / merging Buffer clearing after retrieval n # & B i (t) = ln "(t! t k )!d $ % ' ( k=1 Number of presentations Decay parameter Time since nth presentation Cognitive Modeling WS 2012/13 21 Cognitive Modeling WS 2012/13 22 The noise term is computed by a logistic distribution and is a sum of the noise in the time of retrieval and the permanent noise associated with a chunk. Chunks will be retrieved only if their activation is over a threshold. Because of the noise, there is only a probability that any chunk will be above the threshold: Density s =.2 s =.5! 2 = " 2 3 s2 recall probability i = 1! "A i s 1+ e Retrieval threshold Activation Value variance Cognitive Modeling WS 2012/13 23 Cognitive Modeling WS 2012/13 24 Probability of Retrieval Activation Central component responsible for the coordination in the behaviour of all other modules. Detect the patterns that appear in the buffers and decide what to do next to achieve coherent behaviour. Serial production rule execution: only one can be selected. How does it choose? The utility value! Value of the objective Expected cost of achieving that objective The expected probability that production i firing will lead to a successful completion of the current objective 4

5 Cognitive Modeling WS 2012/13 25 Cognitive Modeling WS 2012/13 26 As with activations for chunks, there is also a threshold which specifies the minimum utility necessary for a production to fire. Probability of executing a production if there are more than one which currently match: Parameters learning Production parameters will change as experience is gathered about the relative costs of different methods and their relative probabilities of success. P(rule i ) = i eu 2s " k e U k 2s Cognitive Modeling WS 2012/13 27 Cognitive Modeling WS 2012/13 28 Some examples Some examples Cognitive tutors for Mathematics. Platform for research on learning and cognitive modeling as part of the Pittsburgh Science of Learning Center. Modeling complex real-world tasks The Dynamic Task: The Anti-Air Warfare Coordinator (AAWC) Integrating brain imaging data Tracking Multiple Buffers in an fmri Study Cognitive Modeling WS 2012/13 29 References An Integrated Theory of the Mind By: John R. Anderson, Daniel Bothell, Scott Douglass, Christian Lebiere and Yulin Qin (Carnegie Mellon University); Michael D. Byrne (Rice University) Diverse materials from ACT-R website: 5