Quantum Models of Human Judgments Jennifer S. Trueblood Vanderbilt University Outline 1. Similarity Judgments 2. Order Effects in Inference 3. Causal Reasoning Similarity Judgments Similarity-Distance Hypothesis Similarity is a decreasing function of distance Distance Axioms 1. D(X,Y) > 0, X ≠ Y 2. D(X,Y) = 0, X = Y 3. D(X,Y) = D(Y,X) symmetry 4. D(X,Y) + D(Y,Z) > D(X,Z) triangle inequality Asymmetry Finding (Tversky, 1977) • How similar is Red China to North Korea? • Sim(C,K) • How similar is North Korea to Red China? • Sim(K,C) • Sim(K,C) > Sim(C,K) Tversky’s Similarity Feature Model • Based on differential weighting of the common and distinctive features • Weights are free parameters and alternative values lead to violations of symmetry in the observed or opposite directions !"#"$%&"'( !, ! = !" ! ∩ ! − !" ! − ! − !"(! − !)! Quantum Model of Similarity sim(A, B) = ||PB PA | i|| 2 Pothos, E., Busemeyer, J. R., & Trueblood, J. S. (2013). A quantum geometric model of similarity. Psych Review. Ko re a Ko re a A quantum account of asymmetry C hina C hina sim(K, C) = ||PC PK | i|| = ||PC | K i|| 2 ||PK | i|| 2 ||PK | i|| = ||PC | i|| ||PC | K i|| 2 > ||PK | 2 C i|| 2 2 sim(C, K) = ||PK PC | i|| = ||PK | C i|| 2 ||PC | i|| 2 2 State vector is assumed to be “neutral” 2 Projection to a subspace of larger dimensionality will preserve more of the amplitude of the state vector Triangle Inequality (Tversky, 1977) • R = Russia, J = Jamaica, C = Cuba D(R,J) < D(R,C) + D(C, J) • Sim(R,J) > Sim(R, C) + Sim(C,J) Findings 1. Sim(R,C) is large (politically) 2. Sim(C,J) is large (geography) 3. Sim(R,J) is small • How can Sim(R,J) be so small when Sim(R,C) and Sim(C,J) are both large? n Cu Ru ssia ba C ar ib b ea C om m unist Quantum Account of the Triangle Inequality Ja m a c i a Sim(R, C) / ||PC | Sim(C, J) / ||PJ | R i|| 2 2 = cos ✓RC = 0.79 2 2 i|| = cos ✓CJ = 0.79 C Not c om m unist No Sim(R, J) / ||PJ | t C ar ib b ea n 2 2 i|| = cos (✓RC + ✓CJ ) = 0.33 R Order Effects in Inference Order Effects ≠ Order Effects in Inference Order effects in jury decision-making: P(guilt | prosecution, defense) ≠ P(guilty | defense, prosecution) The events in simple classical probability models do not contain order information and they commute: P (G|P, D) = P (G|D, P ) To account for order effects, classical probability models need to introduce order information: P (G|P \ D \ O1 ) 6= P (G|P \ D \ O2 ) A Quantum Explanation of Order Effects • Quantum probability theory provides a natural way to model order effects • Two key principles: • Compatibility • Unicity Compatibility Compatible events 1.Two events can be realized simultaneously 2.There is no order information Compatibility Compatible events 1.Two events can be realized simultaneously 2.There is no order information Incompatible events 3.Two events cannot be realized simultaneously 4.Events are processed sequentially Compatibility Compatible events } 1.Two events can be realized simultaneously 2.There is no order information Incompatible events 3.Two events cannot be realized simultaneously 4.Events are processed sequentially Classic Probability } Quantum Probability Unicity • Classical probability theory obeys the principle of unicity - there is a single space that provides a complete and exhaustive description of all events • Quantum probability theory allows for multiple sample spaces • Incompatible events are represented by separate sample spaces that are pasted together in a coherent way Example Voting Event Ideology Event: 1. democrat (D) 1. liberal (L) 2. republican (R) 2. conservative (C) 3. independent (I) 3. moderate (M) Vector Space For Incompatible Events Represented by two basis for the same 3 dimensional vector space D • L C R Voting Basis: L = liberal R = republican C = conservative I = independent M • Ideology Basis: D = democrat V = {|Di, |Ri, |Ii} I • M = moderate Id = {|Li, |Ci, |M i} Ideology Basis is a unitary transformation of the Voting Basis: Id = {U |Di, U |Ri, U |Ii} What if Voting and Ideology are Compatible? Classical probability representation p(L) p(C) p(M) p(D) p(D ∩ L) p(D ∩ C) p(D ∩ M) p(R) p(R ∩ L) p(R ∩ C) p(R ∩ M) p(I) p(I ∩ L) p(I ∩ C) p(I ∩ M) Large nine dimensional sample space Multiple Sample Spaces • Quantum probability does not require probabilities to be assigned to all joint events • Incompatible events result in a low dimensional vector space • Quantum probability provides a simple and efficient way to evaluate events within human processing capabilities When are events Compatible versus Incompatible? It is hypothesized, that incompatible representations are adopted when 1. situations are uncertain and individuals do not have a wealth of past experience 2. information is provided by different sources with different points of view Experiment 1: Order Effects in Criminal Inference 291 participants read eight fictitious criminal cases and reported the likelihood of the defendant’s guilt (between 0 and 1): 1. Before reading the prosecution or defense 2. After reading either the prosecution or defense 3. After reading the remaining case Experiment 1: Order Effects in Criminal Inference 291 participants read eight fictitious criminal cases and reported the likelihood of the defendant’s guilt (between 0 and 1): 1. Before reading the prosecution or defense 2. After reading either the prosecution or defense 3. After reading the remaining case Two strength levels for each case: strong (S) and weak (W) Eight total sequential judgments (2 cases x 2 orders x 2 strengths) Example People v. Robins Indictment: The defendant Janice Robins is charged with stealing a motor vehicle. Facts: On the night of June 10th, a blue Oldsmobile was stolen from the Quick Sell car lot. The defendant was arrested the following day aFer the police received an anonymous Gp. Example People v. Robins Indictment: The defendant Janice Robins is charged with stealing a motor vehicle. Here is a summary of the prosecuGon’s case: •Security cameras at the Quick Sell car lot have footage of a woman matching Robin’s descripGon driving the blue Oldsmobile from the lot on the night of June 10th. Example People v. Robins Indictment: The defendant Janice Robins is charged with stealing a motor vehicle. Here is a summary of the prosecuGon’s case: •Security cameras at the Quick Sell car lot have footage of a woman matching Robin’s descripGon driving the blue Oldsmobile from the lot on the night of June 10th. •During the day on June 10th, Robins had come to the Quick Sell car lot and had talked to Vincent Brown, the owner, about buying the blue Oldsmobile but leF without purchasing the car. •The car was found outside of the Dollar General. Robins is an employee of the Dollar General. Example People v. Robins Indictment: The defendant Janice Robins is charged with stealing a motor vehicle. Here is a summary of the defense’s case: • Robins’ roommate, Beth Stall, was with Robins at home on the night of June 10th. Stall claims that Robins never leF their home. Example People v. Robins Indictment: The defendant Janice Robins is charged with stealing a motor vehicle. Here is a summary of the defense’s case: • Robins’ roommate, Beth Stall, was with Robins at home on the night of June 10th. Stall claims that Robins never leF their home. • Robins recently inherited a large sum of money. While interested in acquiring a new car, she has no reason to steal one. • Robins has no criminal convicGons. Exp. 1 Results SD versus SP SD versus WP 1 SP,SD SD,SP 0.8 Probability of Guilt Probability of Guilt 0.8 1 0.6 0.4 0.2 WP,SD SD,WP 0.6 0.4 0.2 0 Before Either Case After the First Case 0 Before Either Case After Both Cases WD versus SP 1 SP,WD WD,SP 0.8 Probability of Guilt Probability of Guilt After Both Cases WD versus WP 1 0.8 After the First Case 0.6 0.4 0.2 0 Before Either Case WP,WD WD,WP 0.6 0.4 0.2 After the First Case After Both Cases 0 Before Either Case After the First Case After Both Cases Trueblood, J. S. & Busemeyer, J. R. (2011). A quantum probability account of order effects in inference. Cognitive Science, 35, 1518-1552. Modeling Order Effects Two models of order effects: 1. Belief-adjustment model (Hogarth & Einhorn, 1992) • Accounts for order effects by averaging evidence 2. Quantum inference model (Trueblood & Busemeyer, 2011): • Accounts for order effects by transforming a state vector with different sequences of operators for different orderings of information Belief-Adjustment Model The belief-adjustment model assumes individuals update beliefs by a sequence of anchoring-and-adjustment processes: Ck = Ck 1 + wk · (s(xk ) R) 0 ≤ Ck ≤ 1is the degree of belief in the defendant’s guilt after case k s(xk) is the strength of case k R is a reference point 0 ≤ wk ≤ 1 is an adjustment weight for case k Quantum Inference Model Two complementary hypotheses: h1 = guilty and h2 = not guilty The prosecution (P) presents evidence for guilt (e+) The defense (D) presents evidence for innocence (e-) The patterns hi ⋀ ej define a 4-D vector space Jurors consider three points of view: neutral, prosecutor’s, and defense’s Basis vectors for the three points of view |Nij i |Pij i |Dij i Changes in Perspective Unitary transformations relate one point of view to another State Revision Suppose the prosecution presents evidence (e+) favoring guilt 2 3 2 3 2 3 !h1 ^e+ ↵h1 ^e+ ↵h1 ^e+ 6!h1 ^e 7 6↵h1 ^e 7 6 0 7 7 =) 6 7 6 7 =) 6 4!h2 ^e+ 5 4↵h2 ^e+ 5 4↵h2 ^e+ 5 0 !h2 ^e ↵h2 ^e Upn Positive Evidence N eutral P rosecution P rosecution State Revision Suppose the prosecution presents evidence (e+) favoring guilt 2 3 2 3 2 3 !h1 ^e+ ↵h1 ^e+ ↵h1 ^e+ 6!h1 ^e 7 6↵h1 ^e 7 6 0 7 7 =) 6 7 6 7 =) 6 4!h2 ^e+ 5 4↵h2 ^e+ 5 4↵h2 ^e+ 5 0 !h2 ^e ↵h2 ^e Upn Positive Evidence N eutral • P rosecution P rosecution Projection is normalized to ensure that the length of the new state equals one • When the individual is questioned about the probability of guilt, the revised state is projected onto the guilty subspace State Revision Now, suppose the defense presents evidence (e-) favoring innocence 2 3 2 ↵h1 ^e+ 6 0 7 6 6 7 =) 6 4↵h2 ^e+ 5 4 0 Udp P rosecution h1 ^e+ 3 2 7 6 7 =) 6 5 4 h2 ^e+ h1 ^e h2 ^e 0 h1 ^e 0 Negative Evidence Def ense h2 ^e 3 7 7 5 Def ense State Revision Now, suppose the defense presents evidence (e-) favoring innocence 2 3 2 ↵h1 ^e+ 6 0 7 6 6 7 =) 6 4↵h2 ^e+ 5 4 0 Udp P rosecution • h1 ^e+ 3 2 7 6 7 =) 6 5 4 h2 ^e+ h1 ^e h2 ^e 0 h1 ^e 0 Negative Evidence Def ense h2 ^e 3 7 7 5 Def ense Normalize the project and project onto the guilty subspace •A total of 4 parameters are used to define all of the unitary transformations Example Fits Quantum Model: SD versus WP 1 0.8 0.8 0.6 0.4 SP,SD (data) SD,SP (data) SP,SD (QI) SD,SP (QI) 0.2 0 Before Either Case After the First Case Probability of Guilt Probability of Guilt Quantum Model: SD versus SP 1 0.6 0.4 0.2 0 Before Either Case After Both Cases 0.8 0.8 0.4 SP,SD (data) SD,SP (data) SP,SD (Avg) SD,SP (Avg) 0.2 0 Before Either Case After the First Case After Both Cases After the First Case After Both Cases Averaging Model: SD versus WP 1 Probability of Guilt Probability of Guilt Averaging Model: SD versus SP 1 0.6 WP,SD (data) SD,WP (data) WP,SD (QI) SD,WP (QI) 0.6 0.4 WP,SD (data) SD,WP (data) WP,SD (Avg) SD,WP (Avg) 0.2 0 Before Either Case After the First Case After Both Cases Fits to the Data Two models (averaging and quantum) were fit to twelve data points for eight crime scenarios Both models have the same number of parameters (i.e., 4) R2 values for the models: 1. Averaging: R2 = 0.76 2. Quantum: R2 = 0.98 Coffee Break Causal Reasoning Causal Reasoning • Causal Learning • Causal reasoning from verbal descripGons l a c ity i s ioln Learning the effect of a cs hemical b 2y000) a & Sahanks, DNA mutaGons (Lober l C ob or e r Causal r easoning f rom h P T staGsGcal descripGons • Imagine you exercise hard in April. How likely is it that you weight less in May? (Fernbach et al., 2010) Causal Bayes Nets Account for violations of classical probability theory by adding hidden variables (e.g., Rehder, 2014) Z" X" Y" Common%%Cause%%%%%%%% Z" Z" X" Y" W" X" Z" Y" W" X" Y" W" Quantum Approach: A Hierarchy of Mental RepresentaGons • Perhaps(people(use( different(mental( representa2ons(in( different(situa2ons( X Classical Probability Theory Y E Quantum Probability Theory A"Hierarchy"of"Mental"Representa4ons" " " 2"Dimensional" Representa4on" " • All"variables"are" incompa4ble" " " 4"Dimensional" Representa4on" " • Single"cause@effect" rela4onships"are" compa4ble" " " 8"Dimensional" Representa4on" " • All"variables"are" compa4ble" • Classical"probability" model" 2D Model y0# x0# e0# X" | i y1# x1# e1# Three%free%parameters:% % ✓x Rota/on%for%x%basis% ✓y Rota/on%for%y%basis% ↵1 Loca/on%of%state%vector% Y" E" 4D Model 4D vector space with two bases: $ x1 x1 x0 x0 e1$ e0$ e1$ e0$ $ y1 y0 y0 y1 e1$ e0$ e1$ e0$ 6 free parameters: -­‐Three rotaGon parameters ✓1 , ✓2 , ✓3 -­‐Three state vector locaGon parameters ↵1 , ↵2 , ↵3 ! ! 8D Model 8D vector space with a single bases: x1 y1 e1,& &x1 y1 e0,& &x1 y0 e1,& &x1 y0 e0& x0 y1 e1,& &x0 y0 e0,& &x0 y0 e1,& &x0 y1 e0& No rotaGons 7 free parameters to define the locaGon of the state vector ! Causal Power • Parameteriza)on,of,the,8D,model,by,causal,power,theory, (Cheng,,1997), • Each,cause,has,a,power,parameter,w" • Also,allows,for,unknown,alterna)ve,causes,wa" • Condi)onal,probabili)es,are,calculated,by,a,“noisyGor”, equa)on:, p(e1 |xj , yk ) = 1 (1 wx )j (1 wy )k (1 wa ). , • Joint,probabili)es,are,calculated,by,combining,condi)onal, probabili)es,with,prior,probabili)es,of,the,causes:, p(e1 , x1 , y1 ) = [1 (1 wx )(1 wy )(1 wa )]p(x1 )p(y1 ) • Associate)coordinates)of)the)state)vector)with)the)joint) probabili5es)from)causal)power) • 5)free)parameters:) – Three)power)parameters:)wx#,#wy#,#wa# – Two)prior)probabili5es)pr(x1))and)pr(y1)) Quantum'Model'Predic1ons' Three'key'predic1ons:' ' 1. Order'effects' – 2D'and'4D'models' 2. Reciprocity'(Inverse'fallacy)' – 2D'model' 3. Memoryless'effects' – 2D'model' Experiment 1: Order Effects Order Effects • The$sequen)al$processing$of$events$in$quantum$ probability$theory$naturally$gives$rise$to$order$ effects.$ Pr(E|X, Y) ≠ Pr(E|Y, X) Experiment 1 • Causal&reasoning&about&a&novel&biological& category,&Lake&Victoria&Shrimp&(Rehder,&2003)& • Three&binary&features:& High or low ACh neurotransmitter Accelerated or decelerated sleep cycle Normal or high body weight Feature RelaGonships Two causal relaGonships: High%ACh%!%High%Body%Weight% Low%ACh%!%Normal%Body%Weight% Accelerated Sleep Cycle ! High Body Weight Decelerated Sleep Cycle ! Normal Body Weight Sequential Judgments Pr(high'body'weight)' Pr(high body weight | high ACh) Pr(high body weight | high ACh, decelerated sleep cycle) Results (N = 94) p(e = 1| x = 1, y = 1) vs p(e=1| y = 1, x = 1) 80 x = 1, y = 1 y = 1, x = 1 60 40 20 0 Prior 100 Probability of Effect Probability of Effect 100 p(e=1| x = 1, y = 0) vs p(e=1| y = 0, x = 1) 80 60 40 20 0 After First Cause After Both Causes p(e=1| x = 0, y = 1) vs p(e=1| y = 1, x = 0) 80 x = 0, y = 1 y = 1, x = 0 60 40 20 0 Prior After First Cause After Both Causes Prior After First Cause After Both Causes p(e=1| x = 0, y = 0) vs p(e=1| y = 0, x = 0) 100 Probability of Effect Probability of Effect 100 x = 1, y = 0 y = 0, x = 1 80 x = 0, y = 0 y = 0, x = 0 60 40 20 0 Prior After First Cause After Both Causes Modeling Experiment 1 Fit$each$model$to$40$mean$judgments$(20$mean$judgments$per$condi8on)$ Experiment 1: 2−Dimensional Model Experiment 1: 8−Dimensional Model (Unrestricted) 1 1 0.9 0.8 0.8 0.7 0.7 0.6 0.6 Model Model 0.9 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 BIC = -185.88 0.1 0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 BIC = -171.70 0.1 1 0 0.1 0.2 0.3 0.4 Data 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.4 0.3 0.3 0 BIC = -180.90 0 0.1 0.2 0.3 0.4 0.5 Data 0.6 0.7 0.8 0.7 0.8 0.9 1 0.5 0.4 0.1 0.6 Experiment 1: 8−Dimensional Model (Causal Power) 1 Model Model Experiment 1: 4−Dimensional Model 1 0.2 0.5 Data 0.9 1 BIC = -174.97 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 Data 0.6 0.7 0.8 0.9 1 Experiment 2: Invariances of the 2D model Two Invariances Reciprocity:+ Pr(z1+|+x1)+=+Pr(x1+|+z1)++ + Memoryless+effect:+ Pr(x1+|+y1)+=+Pr(x1+|+z1,+y1)+ + + Y" X" Z" Results Exp. 2 • Within&subjects&design&(N&=&58)& • S6muli,&instruc6ons,&and&learning&stage& iden6cal&to&Experiment&1& Probabili(es+ Reciprocity+ Memory+less+ Bayes+Factor+ Pr(x1+|+z1)+ Pr(z1+|+x1)+ 6.41+ Pr(y1+|+z1)+ Pr(z1+|+y1)+ 7.43+ Pr(x1+|+z1)+ Pr(x1+|z1,+y1)+ 2.13+ Pr(x1+|+z0)+ Pr(x1+|z1,+y0)+ 2.96+ Pr(y1+|+z1)+ Pr(y1+|z1,+x1)+ 1.37+ Pr(y1+|+z0)+ Pr(y1+|z1,+x0)+ 6.88+ Conclusions The 2-­‐D model suggests that only one piece of informaGon is processed at a Gme A mental representaGon without joint events might be a simple and more efficient way to evaluate informaGon Related findings: -­‐“singularity principle” of hypotheGcal thinking (Evans, 2006) -­‐ “structurally local” causal reasoning (Fernbach & Sloman, 2009) Thank You What’s coming next… 11:30-12:00 pm Quantum dynamics Part 1 (Peter Kvam) 1:00-2:00 pm Quantum dynamics Part 2 (Peter Kvam) 2:00-2:30 pm Advanced tools for building quantum models part 1 (James Yearsley) 2:30-3:00 pm coffee 3:00-3:45 pm Advanced tools for building quantum models part 2 (James Yearsley) 3:45-4:00 pm Wrap up and discussion (James, Jennifer, and Peter)