A Computational Model of A Computational Model of Routine - PowerPoint PPT Presentation

A Computational Model of A Computational Model of Routine Procedural Memory Routine Procedural Memory Frank Tamborello Frank Tamborello Department of Psychology Department of Psychology Rice University Rice University Houston, TX 77005 Houston, TX 77005 tambo@rice.edu tambo@rice.edu http://chil.rice.edu/ http://chil.rice.edu/

Overview  Very Brief Introduction  Two Experiments, Very Briefly  ACT-R Model 2

Context 3

Contention Scheduling Model (CSM) HORIZONTAL VERTICAL THREADS PROCESSING THREADS SCHEMAS Sensory Information SENSORY- External & TRIGGER PERCEPTUAL Internal DATA STRUCTURES Actions BASE PSYCHOLOGICAL PROCESSING STRUCTURES Motivational Attentional resources influence on add to or decrease activation activation values 4

Simple Recurrent Network (SRN) 5

GOMS  GOAL: EDIT-MANUSCRIPT  GOAL: EDIT-SUBTASK repeat until no more subtasks  GOAL: ACQUIRE-SUBTASK ✦ GET-NEXT-PAGE if at end of manuscript page ✦ GET-NEXT-TASK  GOAL: EXECUTE-SUBTASK ✦ GOAL: LOCATE-LINE – [select: USE-QUOTED-STRING-METHOD – USE-LINEFEED-METHOD] ✦ GOAL: MODIFY-TEXT – [select: USE-SUBSTITUTE-COMMAND – USE-MODIFY-COMMAND] – VERIFY-EDIT 6

ACT-R  Inputs:  Knowledge  IF-THEN rules (termed “productions”)  Declarative knowledge (“chunks”)  Subsymbolic parameters  Simulated task environment/world  Output: Time-stamped behavior sequence 7

Experiment Overview  Task is a routine procedure  Subjects trained approximately one week before  Concurrent working memory task given 8

2 1 ,7 10 6 9, 11 8 3 5 12 4

2 1 , 4 10 3 9, 11 8 5 7 12 6

Mean Total Error Rate Mean T otal Error Rate 0.02 0.04 0.06 0.08 0.12 0.14 0.16 0.1 0 static, intervening subtask procedure change, pre-change Experiment 1 Condition Experiment 1 Condition procedure change, post-change non-intervening semantic control

2 1 , 4 10 3 9, 11 8 5 7 12 6

9, 11 5 6 8 7 1 4 2 3 10 12

9, 11 5 6 8 7 4 2 1 3 10 12

Mean Total Error Rate Mean T otal Error Rate 0.02 0.04 0.06 0.08 0.12 0.14 0.16 0.1 0 static, different-scanner static, same-scanner Experiment 2 Condition Experiment 2 Condition change procedure, pre-change change procedure, post-change static subtask reordering

The Model  Model Goal: Simulate error rates across conditions and trial types  4 conditions  14 trial types total  not just error generation, but also recovery  Highest human SEM error rate = 0.0415  model should do no worse across the board 16

Basic Model Functioning Retrieve Find Move Act Verify Specify next NO Error? action Error Recovery YES Try again to retrieve the action Retrieve another action 17

Procedure Change NO Current step = Retrieve Find Move Act flagged step? YES Verify Specify next NO Error? Retrieve New action Procedure's Step Error Recovery YES Try again to retrieve the action Retrieve another action 18

Mean Error Rate 0.02 0.04 0.06 0.08 0.12 0.14 0.16 0.1 0 static, intervening subtask Experiment 1 Condition procedure change, pre-change procedure change, post-change non-intervening model humans semantic control

Mean Error Rate 0.02 0.04 0.06 0.08 0.12 0.14 0.16 0.1 0 static, different-scanner Jammer, Experiment 2 Condition static, same-scanner model humans change procedure, pre-change change procedure, post-change static subtask reordering

Mean Error Rate 0.02 0.04 0.06 0.08 0.12 0.14 0.16 0.1 0 static, different-scanner Transporter, Experiment 2 Condition static, same-scanner model humans change procedure static subtask reordering

Model Discussion  Discrete, hierarchical goals  governed basic behavior  enabled extensible behavior 22

Basic Model Functioning Retrieve Find Move Act Verify Specify next NO Error? action Error Recovery YES Try again to retrieve the action Retrieve another action 23

Procedure Change NO Current step = Retrieve Find Move Act flagged step? YES Verify Specify next NO Error? Retrieve New action Procedure's Step Error Recovery YES Try again to retrieve the action Retrieve another action 24

2 1, 7 10 6 9, 11 8 3 5 12 4

Model Discussion  No quantitative, multi-condition error models in literature  Same model mechanisms across  4 between-subjects conditions  14 trial types 26

Future Work  Extend model  Step-level error  Step completion time  Model training, too 27

General Discussion  Hierarchical, discrete goal representation matters  …for changing circumstances  …for error recovery  …like CSM  Botvinick & Plaut ’ s connectionist model too narrow  No postcompletion errors  No error recovery  No adaptation of old procedures to new circumstances 28

Acknowledgments  Mike  Carissa Chang & Adam Purtee  Rick Cooper & Jay McClelland  Kristen 29

Thank you!  Questions? 30