Bounded Rationality: The Role of Psychological Heuristics in OR Konstantinos Katsikopoulos Max Planck Institute for Human Development Center for Adaptive Behavior and Cognition July 12–15, 2015 EURO 2015
Psychological heuristics are formal models for making decisions that (i) rely heavily on core psychological capacities, (ii) do not necessarily use all available information, and process the information they use by simple computations, and (iii) are easy to understand, apply and explain. Katsikopoulos (2011), Decision Analysis
900 £ monthly rent 850 £ monthly rent 5 km from city center 25 km from city center no access to garden access to garden Hogarth and Karelaia (2005), Management Science
Have psychological heuristics actually been applied in the field? Did it work? What were the challenges? Do practitioners care?
In 1,060 incidents in NATO checkpoints in Afghanistan, there were 7 suicide attacks and 204 civilian casualties. Can we reduce them? With standard models or psychological heuristics?
Keller and Katsikopoulos (in press), EJOR Fast and frugal tree
Keller and Katsikopoulos (in press), EJOR
Psychological heuristics and soft/hard OR: Conceptual connections
Katsikopoulos (submitted), Handbook of Behavioural OR
Can we flag banks at the risk of failing? To find out, we tested logistic regression and fast and frugal trees in a database of 118 global banks with $100b at the end of 2006, of which 43 failed and 75 survived the crisis.
Aikman et al (2014), Bank of England Working Paper Fast and frugal tree
Aikman et al (2014) Bank of England Working Paper
Systematic studies of psychological heuristics: When do they outperform standard models and when not?
Psychological heuristics have been applied to relevant problems of multi-attribute choice, classification and forecasting. In some problems, it is not clear how to apply standard models of statistics, computer science or hard OR. Can psychological heuristics scale up to more complex problems such as strategic problems with unclear objectives and multiple disagreeing stakeholders (discussed in French et al, 2009)?
Back-up slides
1. Not large performance differences. 2. Simple heuristics are superior in prediction. 3. Each model can outperform the other.
Human Expertise Multiple-occupants cue Leverage ratio Recognition Hiatus Scratch cue Other single cues
Environmental Statistics Small size of training set Predictability of criterion (given the cues) Redundancy of cues (due to the criterion) Lower variance Environmental Structure Non-compensatory binary cues (weights or validities) Error cancellation (more than two alternatives with binary cues) Dominant alternatives (simple or cumulative) Competitive bias
The bias-variance tradeoff
Under risk, an effort-accuracy tradeoff holds. But not necessarily so under uncertainty: prediction error = (bias) 2 + variance + noise Bias is the mean difference between the estimated function and the true function. Variance is the variance around the mean estimated function. Gigerenzer and Brighton (2009), TopiCS
Cumulative dominance
Option A cumulatively dominates option B whenever: Σ k i a k (A) Σ k i a k (B). e.g., a 1 (A) = 1, a 2 (A) = 0, a 3 (A) = 1 a 1 (B) = 0, a 2 (B) = 1, a 3 (B) = 0 Kirkwood and Sarin (1985), Operations Research
For additive utility functions, U(A) = Σ i i a i (A), cumulative dominance characterizes the optimality of lexicographic heuristics. For multi-linear utility functions, U(A) = Σ i i a i (A) + Σ i Σ j > i i, j a i (A)a j (A) + … + i, j,…,k a i (A)a j (A)… a k (A), cumulative dominance leads to the optimality of lexicographic heuristics. Baucells, Carasco and Hogarth (2008) Baucells et al (2008), Operations Research Katsikopoulos et al (2014), EURO J. Decision Processes
Şimşek (2013), NIPS
Şimşek and Buckmann (2015)