Improving Group Decision Making Under Uncertain Circumstances: Applications in Defense Acquisition Dennis Goldenson & Bob Stoddard (SEI) Ricardo Valerdi (University of Arizona) COCOMO 2013 23 October 2013 © 2013 Carnegie Mellon University Information Flow for Early Lifecycle Estimation (QUELCE) Proposed Material Solution & Analysis of Alternatives Expert Judgements Information from Analogous Programs/Systems Program Execution Change Drivers System Characteristics Trade-offs •KPP selection •Systems Design •Sustainment issues ... Operational Capability Trade-offs •Mission / CONOPS •Capability Based Analysis ... Technology Development Strategy •Production Quantity •Acquisition Mgt •Scope definition/responsibility •Contract Award Driver States & Probabilities Plans, Specifications, Assessments Probabilistic Modeling (BBN) & Monte Carlo Simulation Cost Estimates •analogy •parametric •engineering •CERs Program Execution Scenarios with conditional probabilities of drivers/states Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 3 Expert Judgment: Dependency Structure Matrix Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 4 Issues with Expert Judgment Most people are significantly overconfident and overoptimistic in their judgment! Calibrated = more realistic size and wider range to reflect true expert uncertainty An Estimate of SW Size Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 5 5 Studies Confirm Expert Judgment Issues Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 6 6 Cost Estimation Research Previous calibration research Current research in progress Future research & applications Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 7 Calibration Training A series of training exercises • Typically 3 or 4 in sequence Each exercise includes: • A battery of factual questions – Asking for upper and lower bounds within which people are 90 percent certain the correct answer lies – Sometimes true false questions where people provide their confidence in their answers • Brief reviews of the correct answers – Group discussions of why the participants answered as they did – Guidance with heuristics about ways to explicitly consider interdependencies among related factors ... that might affect the basis of one’s best judgments under uncertain circumstances Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 8 Example Open Source Software Questions Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 9 Example Open Source Reference Points Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 10 Example Open Source Reference Points Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 11 A Study of Accuracy versus Precision Which would you rather have? • Someone whose recognized bounds of uncertainty include the correct answer... • Someone who’s a little overconfident but is closer to being accurate... Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 12 Relative Accuracy Improves Experiments confirm that expert judgment can be calibrated. Domain Specific Tests n=29 Generic Tests N=14 Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 13 Training Leads to Better Recognition of Uncertainty Generic Tests Domain Specific Tests Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 14 Experts Improved with Training Test 2: Accurate & imprecise Test 1: Inaccurate & imprecise Test 3: Accurate & Precise Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 15 Cost Estimation Research Previous calibration research Current research in progress Future research & applications Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 16 “Change Drivers” Explain Program Execution Categories of unanticipated change events that often occur in MDAPs over the acquisition lifecycle: • Often a result of previous changes • Leading to subsequent changes • Or affecting program outcomes (which themselves can be drivers of further change. • Status of MDAP activities that are proceeding as planned are not change drivers. Intended use • To enable DoD domain specific expert judgment training – Initially in QUELCE workshops • Other uses may be possible if we are successful in populating a larger DoD domain-specific reference point repository, e.g.: – “Deep dives” earlier in pre-Milestone A – Program planning & risk analysis throughout the Acquisition lifecycle Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 17 Domain Reference Points Aid Judgment “There is a 90% probability that MDAPs with certain characteristics will experience off nominal change drivers A, B and C.” “When change driver A goes off nominal, there is a 75% probability change driver B will go off nominal” “When change drivers A, B, and C go off nominal, there is a 90% probability that change driver D will go off nominal.” “When specific change drivers go off-nominal, specified impacts have occurred.” “When specific change drivers go off-nominal, other change drivers are influenced or impacts felt within a certain amount of calendar time.” Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 18 A Reference Point Repository for DoD Categorizing & tagging textual information about change events • From program documents such as SAR & DAES • Identify DoD domain specific reference points mapped to QUELCE change drivers Joining the tags & text excerpts with existing data • MDAP domain characteristics • Program performance outcomes, e.g., cost, schedule &scope of deliverables Using the categories & text excerpts: • To assist judgments by QUELCE workshop teams based on experiences in analogous programs • For use in individual expert calibration experiments & group resolution of differences among team members If we’re successful: Also used to support other activities • Both earlier in Milestone A & throughout the program lifecycle Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 19 Cost Estimation Research Previous calibration research Current research in progress Future research & applications Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 20 What’s Next for Expert Judgment Research? A focus on DoD domain-specific questions & reference points Seed a queryable reference point repository with DoD data Shift our focus to experiments on resolution of differences among members of expert groups • Quantify benefit of access to domain reference points • Comparing algorithmic & group consensus methods with respect to accuracy, recognition of uncertainty, & time required to resolve differences among team members Upgrade our existing software support: • To capture individual judgments & eventually resolve differences without the need for face-to-face meetings Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 21 Leveraging the Delphi Planning Process Given historical work • Wideband Delphi applied to cost estimation enabling discussion & a broader communications channel to produce more accurate results (Boehm 1981) • Recent research in software project estimation shows that estimates that benefit from group discussion tend to be more accurate (Cohn, 1997; Moløkken & Jørgensen, 2004). We will research improved group decision-making judgment • Leverage expertise to forecast uncertainties related to costs and risks of program execution • Revisit conventional Delphi discouragement of discussion between rounds, introducing discussion of domain reference points Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 22 Additional Considerations in Judgment Experiments Heuristics such as anchoring & adjustment, overconfidence, blind spot bias, and others commonly bias individual experts’ judgments • An individual estimator may first make a “best estimate” of duration for a program element ... then adjust it to form long-duration and short-duration estimates giving a range of likely outcomes • Such adjustments are commonly known to be too small (Fischhoff, 1994) • Resulting in too-tight range estimates & hugely over-frequent 1% and 5% tail occurrences • However, explicit prompting of the estimator’s imagination can substantially reduce this tightness (Connolly & Deane, 1997) Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 23 Summary The eventual target is to apply these & related group reconciliation methods in our current research on QUELCE • QUELCE works by codifying expert judgment for cost estimates prior to Milestone A (Ferguson et al., 2011 ) • However improving group decision making is equally important for program planning and risk analysis throughout the lifecycle. We will validate & enhance our previous research on calibrating individual judgment (Goldenson & Stoddard, 2013) by: • Developing DoD domain-specific questions for a series of test batteries & associated training exercises • Investigating the value of DoD domain-specific reference points that provide more detailed contextual background about programs analogous to the programs being considered in calibration test questions We welcome collaborators for the expert judgment experiments! Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 24 Contact Information Dennis R. Goldenson Senior Member of the Technical Staff Software Engineering Institute Telephone: +1 412-268-8506 Email: dg@sei.cmu.edu U.S. Mail Software Engineering Institute Customer Relations 4500 Fifth Avenue Pittsburgh, PA 15213-2612 USA Web www.sei.cmu.edu www.sei.cmu.edu/contact.cfm Customer Relations Email: info@sei.cmu.edu Telephone: +1 412-268-5800 SEI Phone: +1 412-268-5800 SEI Fax: +1 412-268-6257 Improving Group Decision Making Under Uncertain Circumstances COCOMO 2013, 23 October © 2013 Carnegie Mellon University 25