When-to-act: Evidence for Action Bias in a Dynamic Belief-Updating Task Michael Hildebrandt Human-Computer Interaction Group Department of Computer Science University of York Joachim Meyer Department of Industrial Engineering and Management Ben Gurion University DIRC annual workshop Edinburgh, 15-17 March 2005 Time Design Time as designable property of systems When-to-act problem Managing action/judgement [timeliness/accuracy] trade-offs Modelling Bayesian updating, SDT, normative / empirical When-to-act experiment Diagnostic task, belief updating [alarm], action bias When-to-act: Evidence for action bias in a dynamic belief-updating task Theme activities: Time Design workshops CHI’04: Subjective (psychological, social time), representation ECCE’04: Control of dynamic systems, social time HFES’04: Temp. decision, interruption scheduling, case studies Conclusion: From Newtonian view to functional view [Slides, wiki, mailing list: www.timeDsn.net] ‘Newtonian’, descriptive view of time ’Absolute, true and mathematical time, of itself and from its own nature, flows equably and without relation to anything external.’ Descriptive view of time. Problem: ubiquity. Events necessarily happen in time, but this does not make time an interesting or relevant property in design/operation KSLM, Newell’s “bands”, task representation, time-and-motion Time Design – When-to-act Time Design: problem An interdisciplinary – Model – When-to-act view experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Time Design: Functional view of time Time as designable property of systems Psychological time, social time Designing the experience of time Designing the display of temp. inf., supporting temp. awareness Tempo: Increase speed / availability / flexibility; results in fast / dynamic / complex systems; creates need for machine support; creates need to design effective human control of system Designing / appreciating temp. patterns Supporting synchronisation / coordination Design space: Time as property of environment, task, physical system / interface, user behaviour Functional view: Exploring degrees of freedom, trade-offs Time Design – When-to-act Time Design: problem An interdisciplinary – Model – When-to-act view experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Representation, analysis, design From [focus on] temp. description of human behaviour (KSLM) To Human reaction to / use of time Time in HCI / Human Factors: Research domains System Response Time Temporal awareness, temporal reference systems Temporal error; decision lockout Control of dynamic systems; automation, scheduling, control Trust in automation / dynamic systems; long-term automation Time perception & stress; affordance model of dynamic system TA Time Design tutorial Time Design – When-to-act Time Design: problem An interdisciplinary – Model – When-to-act view experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Example: Dynamic Function Scheduling research Misperception / misjudgement of when-to-act Misperception of time available / time required Duration neglect in assessing control strategies Inappropriate / rigid Function Scheduling Benefit of temporal awareness and anticipative control behaviour Multiple and conflicting temporal reference systems Scheduling and allocation based on dynamic value functions Late and premature action; pacing and decision lock-out Effect of time pressure on automation use and trust Human scheduling performance; sched. as operation strategy TA Dynamic Function Scheduling Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task When-to-act problems: Simple cases Event-driven, self-paced, periodicity, predictability, triggers Simple heuristics: The sooner, the better When-to-act dilemmas Motivation to act early [reduce uncertainty, increase prob. of execution] Motivation to defer action [improve solution, gather additional inf.] Risk: Action- or judgement-bias leads to premature / late decision Contributing factors: Reliability, predictability, timing, age of info. Temporal uncertainty [reliance on time perception vs. visualisation] When-to-act problems in the real world ATC, medical treatment [A&E, long-term], piloting [V2, pattern], rear-end CW, supervisory control [manufacturing], emergency C&C Time Design – When-to-act Time Design: problem An interdisciplinary – Model – When-to-act view experiment When-to-act: Evidence for action bias in a dynamic belief-updating task When-to-act experiment System-1 information Start of trial Alarm (Sys-2 Decision information) Decision End of trial Investment, rating Trial duration:18s 100 trials, IVs: Timing and reliability of alarm Alarm timing early (5s) late (13s) Alarm high (.9) reliability low (.7) Time Design – When-to-act problem – Model – When-to-act experiment Feedback t When-to-act: Evidence for action bias in a dynamic belief-updating task Modeling and paradigm Diagnostic DM, monitoring & control [Kerstholt] Response to alarms [over-/under-use, compliance/reliance, reliability] Signal detection theory Bayesian belief updating Heuristics & biases Time perception Human Factors approach Normative / engineering model as standard Compare with empirical data Model empirical data (ideal: in same formalism) Identify systematic deviations, biases, behavioural disposition Express these using formal models or design heuristics Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Modeling temporal aspects of the system p(overallSuccess) = p(decisionCorrect) * p(decisionInTime) Alarm timing: Early (mean 5s), late (mean 13s); SD=1.5 Prob. action execution: f(x)=1.0875-x/18*0.675 Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Modeling pre-trial information: SDT Noise distribution (“no problem”): Mean=4, SD=1 Noise distribution (“problem”): Mean=5, SD=1; d’=.5 Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Modeling alarm reliability, combining evidence High reliability: p(miss)=p(FA)=.1; Low: p(miss)=p(FA)=.3 Decision based on pre-trial inf.: p(turb|x)=p(x|turb)*p(turb)/p(x) Decision accuracy after updating with additional information: p(turb|x∩alarm)=p(turb)*p(x|turb)*p(alarm|turb)/p(x∩alarm) Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Modeling strategies p(overallSuccess) = p(decisionCorrect) * p(decisionInTime) Decision based solely on pre-trial information could [should] be taken at the start of the trial, where p(decisionInTime)=1 and p(decisionCorrect)=p(turb|x) Only wait for alarm if p(turb|x) < p(turb|x∩alarm) * p(decisionInTime[alarmTime]) AlarmTime is unknown at the start of the trial, but mean alarm time [5s, 13s] can be used as an approximation Alarm makes biggest improvement at indifference point [4.5] Strategies With a low reliability alarm, never wait for the alarm unless in the early alarm condition with a system-1 value close to the indifference point. With a high reliability alarm and an early alarm, wait for the alarm even for system-1 values around the distribution means (3.7-5.3). With late alarms, only wait if the value is near the indifference point. Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Results: Alarm usage Data suggests participants in the late over-relied on low-reliability alarm and under-relied on high-reliability alarm alarm condition Proportion of trials where participants waited for alarm Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Results: Overall earning Alarm timing and reliability affected overall earning Main effect of reliability more pronounced in late condition Need to compare with normative model [esp. p(turb|x)] Need to analyse investment strategy Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Results: Participant feedback Evidence for action bias “Timeliness is the important characteristic, especially when even some decisions made in 5-6sec are late! So, more important to send the data quickly upon current estimates than wait for accuracy. [After 80 trials] Timeliness is important but accurate data again may not be available in time, need to balance the two. [After 100 trials] I waited long enough, about 5seconds for the alarm to occur. If it did, I almost invariably based my decision on it, otherwise I made my decision on Sensor1 data.” A few late decisions may make participants over-cautious “[after 20 trials] At first I waited to see system 2, but realised it was too slow, so based my decisions on system 1.” Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Results: Overall success (trend) Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Results: Decision timeliness (trend) Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Results: Decision accuracy (trend) Time Design – When-to-act problem – Model – When-to-act experiment When-to-act: Evidence for action bias in a dynamic belief-updating task Summary Identified an important but under-researched temporal reasoning problem that may severely compromise dependability Provided a formal model (SDT + Bayesian updating) against which empirical data can be compared Conducted experiment manipulating alarm timing & reliability Contrary to previous studies, obtained evidence for action bias Feed into Timing book, TA Time Design tutorial, TA DFS Future work More detailed analysis of data [invest. strategy, time perception] Modeling of empirical data Additional experiments investigating alternative w2a scenarios, different motivations for deferring/promoting decisions, levels of uncertainty, visualisation of time Derive design heuristics? Time Design – When-to-act problem – Model – When-to-act experiment