Functional encoding in memory for goals ACT-R workshop August 1999 Erik M. Altmann (altmann@gmu.edu) J. Gregory Trafton (trafton@itd.nrl.navy.mil) Means-ends tasks • Means-ends behavior: – Suspend a goal – Work on subgoals – Resume the goal at an appropriate time • Examples: – Monkey and bananas – Giving a talk – Making photocopies The Tower of Hanoi • The foundational means-ends task – In cognitive science • Understood in terms of the goal stack • Completely understood – Or is it? • Good data (Anderson, Kushmerick, & Lebiere, 1993) The Tower of Hanoi Subgoal 3:B Goal 4:C 1 2 3 3 4 4 B A C A stack model Recall 3:B perfectly, despite lag Stack height Push 3:B 1:B 4:C 3:B 4:C 2:C 3:B 4:C 2:C 3:B 4:C Time 2:C 3:B 4:C 3:B 4:C 1:C 3:B 4:C ... The stack as representation • The typical assumption in task analysis – Implicit in problem behavior graph – Explicit in GPS, GOMS, ... • The standard theory of goal management – In cognitive architectures • ACT-R, Soar – In cognitive models generally • E.g., ACT-PRO, 3CAPS Better Raven, ... The stack as representation • The appeal: – Robust and general – Applies to a wide variety of tasks – Supported by empirical data • At some level of abstraction • The problem: – At best, a high-level simplification – At worst, wrong Goal-selection order • LIFO order not used when not needed – Selection order in arithmetic (VanLehn) • Order depends on context – Display-based problem-solving, situated action, distributed representation – Capture error Pending goals • • • • Displaced by memory load (Just & Carpenter) Decay when not rehearsed (Byrne & Bovair) Intrude when rehearsed (Altmann & Trafton, 1999b) Affected by goal content – Intention superiority (Goschke & Kuhl) • Suggesting that activation affects availability Research approach • Model Tower of Hanoi data without a stack – For goals • Ask how to make up the lost functionality – Domain knowledge – External cues – Existing memory theory • If it suffices, the theory is strengthened • If it fails, then at least we know why Memory as goal store (MAGS) • Memory = encoding + retention + retrieval • Assume passive retention • Assume strategic encoding – Using knowledge of retrieval context • Assume strategic retrieval – Using knowledge to select retrieval cues Analytical framework: Activation • What happens to a goal’s activation over time? • Two kinds of activation (in ACT-R): – Base-level activation from use – Priming from context • Total activation predicts current need – So memory returns the most active element Encoding to resist decay • Strengthen base-level activation • Strength test to say how much is enough – Cognition asking itself, “Got it?” • If yes, stop strengthening and move on • If no, strengthen some more – Test interleaved with strengthening • Strengthen enough but not too much Encoding to resist decay Base-level activation Strength test 2:C, 1:B, 2:C Retrieval threshold Time The strength test • Cognition can anticipate retrieval context – Retrieval cue — “3” for 3:B – Retention interval — 5 to 10 seconds • Anticipations are just knowledge – Represent as cue chunks • Test-retrieve the goal – If test fails, encode some more Focussed retrieval 3:B Test retrieval cue: sink: 3 S disk: 3 from: A to: B blocked: t Encoding context Goal Retrieval cue: 3 disk: 3 from: A to: B blocked: t Retrieval focus Main focus Retrieval context Retrieval production (p retrieve =focus> isa retrieval =goal> isa goal disk =disk to =peg ==> =focus> disk =disk to =peg !pop!) No indexing or chaining Noisy retrieval without partial matching Empirical test • Anderson, Kushmerick, & Lebiere (1993) – Subjects instructed in goal-recursion strategy – Response-time data are from perfect trials • Cognition on those trials most stack-like • Strongest test of the MAGS model Prediction • Encoding a goal is expensive – Not a cost-free push operation – A second or so per goal • Prediction from serial attention model Data Time (sec) Large peaks = Goal encoding 12 10 8 6 4 2 0 Observed (AK&L 93) Simulated (MAGS), R 2 = .99 1 3 5 7 9 11 Move in solution path 13 15 Prediction • People avoid unnecessary retrievals – Retrieval is effortful and error-prone • Use move heuristics when they apply: Don’t-undo IF the just-moved disk was 1, and X is the smaller of the two other top disks, and Y is the larger of the two other top disks, THEN move X on top of Y. Data Valleys = Don’t-undo 12 10 8 6 4 2 0 1 3 5 7 9 11 Move in solution path 13 15 Prediction • Prefer goal retrieval to re-planning • Depends on selecting the right retrieval cue – No perfect pop operation • Cue selection heuristic: Retrieve-uncovered IF the uncovered disk is X, THEN try to retrieve X:? Data Small peaks = goal retrieval 12 10 8 6 4 2 0 1 3 5 7 9 11 Move in solution path 13 15 Five-disk data 16 14 Simulated (MAGS), R2 = .95 12 10 8 6 4 2 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 Move in solution path Parameters • ACT-R defaults: – W = 1.0, F = 1.0, d = 0.5 • Adopted from other models: – Perceptual encoding time = 185 msec (Anderson, Matessa, & Lebiere, 1997) – t = 4.0, s = 0.3 (Altmann & Gray) • No unconstrained parameters Prediction • Retrieval is error prone – E.g., might retrieve 3:C instead of 3:B • From a previous plan or previous trial – Incorrect retrieval starts a garden path Data Length of solution path 60 Predicted (MAGS) Observed (AK&L 93) 50 40 Optimal 30 20 10 Optimal 0 Four disks Five disks MAGS vs. stack model (A&L 98) • Based on declarative memory – Not on a privileged stack • Broader empirical coverage – Detailed account of RT and error – Only ToH model to address both (before today) • Functional encoding and retrieval processes – Specified at ACT-R’s atomic level – Generic — adapted from serial attention (Altmann & Gray, 1999b) Implications • Need a two-high architectural stack – A main focus for problem state – A retrieval focus for concentrating • Main and retrieval focuses are mutually exclusive (Altmann & Trafton, 1999b) – One is reliable – One is predictive Conclusions • Don’t need a goal stack – Anything it can do, MAGS can do better – And without that much more analysis • Don’t want a goal stack – Too easy and too wrong – Masks real goal-management mechanisms Conclusions • 40 years of research on the Tower of Hanoi • Yet retrieve-uncovered is unpublished – Missing from Simon’s perceptual strategies – Missing from Anzai and Simon protocol Conclusions • Why now? – Detailed data – A precise memory theory – Throwing away the goal stack References Model code: hfac.gmu.edu/people/altmann/toh Altmann & Trafton (1999a). Memory for goals: An architectural perspective. Proc. Cog. Sci. 21. Altmann & Trafton (1999b). Memory for goals in means-ends behavior. Manuscript submitted for publication. The encoding process Focussed retrieval with a “sink” 4:C Test-retrieval Test-strength cue: sink: 4 S disk: 4 from: A to: C blocked: t disk: 4 from: A to: C blocked: t Strengthen-goal Test-passes/fails disk: 4 from: A to: C blocked: t Test-fails