Automaticity development and decision making in complex, dynamic tasks Dynamic Decision Making Laboratory www.cmu.edu/DDMLab Social and Decision Sciences Department Carnegie Mellon University Cleotilde Gonzalez Rickey Thomas Polina Vanyukov 1 Complex and dynamic tasks Executing a battle, driving, air traffic controlling, managing of a production plan, piloting, managing inventory in a production chain, etc. • Demand real-time decisions (time constraints) • Demand attentional control • Require multi-tasking: they are composed of multiple and interrelated subtasks • Demand the identification of ‘targets’ defined by multiattributes • Demand multiple and possibly changing responses 2 Automaticity in dynamic, complex tasks • targets and distractors are often inconsistently mapped to stimuli and responses • Often, we bring pre-learned categories and mappings to a task stimulus - category L ------------- letter category - response button --------- click • Are decision makers in dynamic situations operating in controlled processing continuously? 3 Proposed model of automaticity in DDM sub-task Structure (Mapping) sub-task Structure (Mapping) CM/VM CM/VM Cues Categories Responses Cues coupling Categories Responses coupling Goals (Relevancy) Task switching (resource allocation) 4 Experiments • Automaticity develops with consistently mapped stimuli to targets, even when targets move and time is limited (Experiment 1) • The consistency of target to response mapping also determines automaticity development (Experiment 2) • Automaticity of a task component frees-up time and resources for high level decision-making (Experiment 3) • Automaticity develops differently with different degrees of pre-learned categories (Experiment 4) 5 The Radar Task 6 7 General method • Independent variables o stimulus mapping (CM or VM) • CM = Search for Numbers in Letters • VM = Search for Letters in Letters o cognitive load • Memory set size (MSS): Number of possible targets to remember (1 or 4) • frame size (FS): Number of blips present on the screen at a given time (1 or 4) o target present/absent (a target was present 75% of the trials) • Dependent variables o o Accuracy: proportion of correct detections or decision-making responses Time: mean target detection or decision-making time in msec • From 18 to 30 hours of practice, 3 hours per day 6 to 10 days 8 Experiment 1: Consistency of stimuli • Replicate major findings from the dual-process theory (Schneider & Shiffrin, 1977) in a dynamic task • Automaticity is acquired with practice in consistent mapping conditions, and automatic performance is unaffected by workload 9 Experiment 1: Method o CM vs. VM o Cognitive Load Variables • Memory Set Size • Frame Size o Only one possible response: pressing spacebar when target is detected 10 Experiment 1: Accuracy 1 CM VM CM 0.9 Accuracy VM 0.8 0.7 0.6 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 11 Experiment 1: Detect Time 1600 1400 1200 Msec VM 1000 VM CM 800 CM 600 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 12 Experiment 1: Summary • Radar’s manipulations of cognitive load interact with stimulus mapping in ways that parallel Schneider & Shiffrin’s results • Automaticity develops with extended practice and consistently mapped stimuli even when targets move and time is limited • Radar task can be used to study automaticity in dynamic stimulus environments 13 Experiment 2: Response Consistency • There is some evidence that response mapping is not critical for automaticity to develop (Fisk & Schneider, 1984; Kramer, Strayer, & Buckley, 1991) • In complex tasks mapping of targets to responses can be inconsistent o Resulting in large processing costs, even when stimuli are consistently mapped to targets 14 Experiment 2: Method o Only consistently mapped stimuli o Cognitive Load Variables • Memory Set Size • Frame Size o Response consistency varied in four levels 15 Response Mapping Conditions T Mapped to Stimuli Partial Mapping to interface T T Fully Mapped to interface Random Mapping T 16 Experiment 2: Accuracy 1 Accuracy 0.9 0.8 0.7 0.6 Stimulus Full Partial Random 17 Experiment 2 : detect time 1600 1400 Msec 1200 1000 800 600 Stimulus Full Partial Random 18 Experiment 2: Summary • A consistent response reduces processing requirements • Total task consistency (both, consistency of stimuli and consistency of responses) matters o There are processing costs if responses are not consistently mapped, even when stimuli are • Implications o Interface design: interface influences processing of responses • Response selection using track-up vs. north-up displays • Make response selection intuitive • Interface design, decision support tools, training o We can now systematically manipulate Radar to elucidate the effects of automaticity on high-level dynamic decision-making 19 Experiment 3: Automatic detection & highlevel decision making • How would automatic detection of a component help decision-making? • Decision-making component required operators to analyze a sensor array of detected aircraft • Sensor and weapon information changed dynamically 20 Experiment 3: Method • Sensor Reading Task • Determine if Target is Hostile o o o Scan Sensors > 13 (Hostile) < 13 (Non-Hostile) • Press Ignore (5-Key) • Select Response (Weapon Systems) o o o Guns vs. Missiles > 10 Missiles (6-Key) < 10 Guns (4-Key) • Quiet Airspace Report o o No targets detected Click submit report with mouse key 21 Experiment 3: Detect Accuracy 1 CM CM VM 0.9 Accuracy VM 0.8 0.7 0.6 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 22 Experiment 3: Decision-making Accuracy 1 Accuracy 0.9 CM CM VM 0.8 0.7 VM 0.6 FS=4 FS=1 MSS = 1 FS=4 FS=1 MSS = 4 23 Experiment 3: Detect Time 1600 Msec 1400 1200 VM 1000 CM VM 800 CM 600 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 24 Experiment 3: Decision-making Time 1600 Msec 1400 1200 1000 VM CM VM CM 800 600 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 25 Experiment 3: Summary • Consistent mapping of targets improved he accuracy of the decision-making of the task • Detect time, detect accuracy, and whole-task performance are sensitive to workload manipulations • Implications o Consistent mapping actually improved whole-task performance by freeing up time for the controlled sensorreading tasks to run to completion o Thus, processing speed-up associated with automatic detection can have a large impact on whole-task performance 26 But…? • Is accuracy of decision-making improved simply because there is more time to process? • Effect of detection on high-level decision-making in the presence of a dual-task 27 Experiment 3b: Method • Secondary tone task: enter count of number of nonstandard tones o Calibrated to standard tone at beginning of session for each participant o Non-standard tones higher/lower pitch than standard 28 Experiment 3b: results • In fact the Radar task performance was the same with and without the tone task! Performance on Tone Task • Detect Time o No Effect of secondary task 0.85 0.83 0.81 • Decision-Making Time o No Effect of secondary task • Decision-Making Accuracy o No Effect of secondary task 0.79 Accuracy • Detect Accuracy o No Effect of secondary task 0.77 CM 0.75 VM 0.73 0.71 0.69 0.67 0.65 MSS = 1 MSS = 4 29 Experiment 3b: Implications • No effect of dual task on RADAR performance • Operators are allocating resources away from tone task to maintain RADAR performance • Implications o Finding supports the hypothesis that consistent mapping improves decision-making performance by freeing up resources for other tasks o Thus, processing speed-up and low resource requirement associated with consistent mapping can have a large impact on performance in complex task 30 Experiment 4: Categorization • Since consistent mapping is the search for numbers in letters, it is possible that load-free processing is due to categorization (Cheng, 1985) • Purpose of this experiment is to establish the presence of load-free processing without categorization 31 Experiment 4: Method • Incorporate memory ensembles where no possible categorization can take place either a priori or with learning • CM vs. VM with tone o o CM = {C, G, H, M, Q, X, Z, R, S} VM = {B, D, F, J, K, N, W, P, L} • Memory ensembles were equated o o Angular {H,M,X,Z,F,K,N,W} vs. Round {C,B,D,G,Q,P,R,J} Beginning {B,C,D,F,G,H,J,K} vs. End {M,N,P,Q,R,W,X,Z} • Cognitive Load Variables o o Memory Set Size (1 or 4) Frame Size (1 or 4) • Indicated detection of target by pressing spacebar o o Detect Performance Detect Response Time 32 Experiment 4: Detect accuracy 1 VM CM CM Accuracy 0.9 VM 0.8 0.7 0.6 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 33 Experiment 4: Decision-making accuracy 1 Accuracy 0.9 VM CM 0.8 CM 0.7 VM 0.6 FS=4 FS=1 MSS = 1 FS=4 FS=1 MSS = 4 34 Experiment 4: Detect time 1600 1400 Msec VM 1200 CM 1000 VM CM 800 600 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 35 Experiment 4: Decision-making time 1600 Msec 1400 1200 1000 VM CM CM VM 800 600 FS=1 FS=4 MSS = 1 FS=1 FS=4 MSS = 4 36 Experiment 4: Implications • Varied mapped performance is more sensitive to load than consistently mapped performance • Individuals performed better in the high-level decision-making component of Radar when stimulus mapping was consistently mapped • Implications o o Categorization is NOT a necessary requirement for automaticity development Consistent stimulus mapping is a necessary condition for the development of automatic detection 37 Summary of accomplishments • Developed Radar, a dynamic simulation where it is possible to study (i.e., to measure) automaticity • In Radar it is possible to elucidate the effects of automaticity on high-level dynamic decision-making • Established the usefulness and applications of the dualprocess theory of automaticity • Deepen our understanding of the implications of automaticity development for practical real-world tasks • Brought together two main theories of automaticity: instancebased theory and dual-process theory 38 Future research • Consistency of mapping and responding is relative to the categories (i.e., similarity) that a user can form • Thus, consistent mapping can lead to automatic responses for high-level decision-making after extended practice 39 Looking towards applications • Test these hypotheses in airport luggage screening • Decide whether to hand search the luggage • There is no consistency but rather just similarity (relative to a ‘knife’ category) 40