Filippo Radicchi 1.5 98 filrad.homelinux.org 1.4 1.3 1.2 filiradi@indiana.edu 105 0 10 β 102 0 10 102 105 Hunting98for numbers: optimal strategies in mental searches Tracking animals Movement patterns in nature albatross Viswanathan GM, et al., Nature 381, 413–415 (1996) marine predators Sims DW, et al., Nature 451, 1098–1102 (2008) spider monkey Fernandez G, et al., Beahv Ecol Sociobiol 55, 223–230 (2004) honey bee Reynolds AM,et al., J Exp Bio 210, 3763-3770 (2007) Movement patterns in nature Lévy flight Humphries NE, et al., Nature 465, 1066–1069 (2010) Shlesinger MF, Zaslavsky GM, Klafter J, Nature 363, 1–37 P (d) d µ Lévy flights are optimal search strategies Viswanathan GM, et al., Nature 401, 911-914 (1999) What about humans? 107 105 104 years ago 102 101 Human mobility patterns currency tracking Brockmann D, et al., Nature 439, 462-465 (2006) Human mobility patterns mobile phone tracking Gonzalez MC, et al., Nature 453, 779-782 (2008) Lévy flights: a universal description for movement patterns of animals and humans from www.christineberrie.com Problems Results for animals are controversial Journal of Animal Ecology Volume 81, Issue 2, pages 432-442, 17 OCT 2011 DOI: 10.1111/j.1365-2656.2011.01914.x http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2656.2011.01914.x/full#f2 Problems 1) Data are aggregated over many “different” individuals. Are we measuring heterogeneity at the individual or at the population level? 2) Three orders of magnitude represent a “luxury”. Is the scaling compatible with a power law? 3) Do modern humans move in the space for search purposes? Lowest unique bid auctions Lowest unique bid auctions game rules 1 2 3 M-2 3 • high-valued goods are put up for auction V = 10 $ A M = 101 $ 101 $ not unique not unique unique and lowest not unique unique but not lowest 1¢ 2¢ 3¢ 4¢ 5¢ 3 bids 2 bids 1 bid M 105 $ • agents independently explore the bid space in search for the winning bid • each bid costs c = 100 $ M-1 2 bids 1 bid the winning bid is the lowest and unmatched bid 103 $ Lowest unique bid auctions data sources bid 8000 6000 A Uniquebidhomes agent 1632 auction 19 4000 change 2000 B 6000 4000 2000 0 100 200 300 400 500 600 700 t Cumulative distribution Patterns of bid space exploration Using the model to fit the data (Q )ji = |i j| [1 (i mj ( ) j)] transition matrix p (b1 , b2 , . . . , bT | ) = (Q )0,b1 (Q )b1 ,b2 (Q )b2 ,b3 · · · (Q )bT fitting data by maximizing the likelihood 1 ,bT Testing the validity of the model 1 P (d) = (Q )ij (d i,j |i (Q )ij (d j|) d i,j goodness of fit |i j|) Probability density Patterns of bid space exploration 2.0 Uniquebidhomes 1.5 1.0 0.5 0.0 0.5 D 1.0 1.5 α 2.0 2.5 Patterns of bid space exploration A Probability density 10 B 0 10 Uniquebidhomes Lowbids Bidmadness 0 positive negative 10 -2 10 -2 10 -4 10 -4 10 -6 10 -6 C10 10 0 10 1 10 2 10 3 10 4 D10 0 10 -2 10 10 T<10 10≤T<40 40≤T<200 T≥200 10 -2 -4 10 -4 -6 10 -6 10 0 10 1 10 2 10 3 10 0 10 1 10 2 10 3 10 4 10 4 0 10 Uniquebidhomes Uniquebidhomes 4 change first bid t≤2 t≤5 t≥T-2 t=T-1 Uniquebidhomes 10 0 10 1 10 2 10 3 Modeling lowest unique bid auctions single auction 0 1 p (i) = i 2 3 M-2 M-1 M M /m ( ) m( ) = j j=1 prob. that the agent bids on value i N = Number of agents T = Number of bids per agent M P ({n}) = N ! k=1 nk [p (k)] nk ! prob. to observe a configuration {n} = (n1 , n2 , . . . , nk , . . . , nM ) M = Max bid value = LF exponent of the population = LF exponent of the agent c = Fee value V = Value of the good Modeling lowest unique bid auctions single auction N 1 u (i) = P (ni = 1) = N p (i) [1 p (i)] prob. that a bid on value i is unique Focus on the agent with strategy l , (v) = p (v) [1 N p (v)] [1 u (k)] k<v prob. that the agent places a unique and lowest bid on value v M w , = M l , (v) v , = vl , (v) v=1 v=1 prob. that the agent wins the auction average value of her winning bid Modeling lowest unique bid auctions multiple auctions G = Total number of auctions g = Total number of wins P , (g) = G g (w , g ) (1 w , G g ) prob. that the agent wins g auctions R , (I |g ) = l , (v1 ) l v1 +v2 +...+vg =I prob. that the sum of her winning bids is I in g wins I) /G , (g) = (gV economic return of the agent r , (v2 ) · · · l , (vg ) Modeling lowest unique bid auctions multiple auctions R (I) = , P , (g) R , (I |g ) g prob. that the sum of her winning bids is I In the limit 1 G g = Gw I= g v , , =w , (V , v , sum of the winning bids average number of wins r = Gw , v economic return of the agent , ) r , >c positive return r , <c negative return Modeling lowest unique bid auctions multiple bids q (1) (i) = i (Q )ji = /m ( ) first bid |i j| q (i) = T (Q )ji q (t 1) (j) s(T ) (i) = 1 j=1 next bids (T ) u , 1 q (t) (i) t=1 probability to visit a given site (i) = N s (T ) (i) 1 (T ) s (i) N 1 (T ) s (i) (T ) u , (v) 1 prob. that the agent makes on value i a unique bid (T ) l , j)] transition matrix M (t) [1 (i mj ( ) (T ) (v) = s (v) 1 (T ) s (v) N 1 k<v prob. that the agent makes on value i a lowest and unique bid Model predictions 0 0 0 1.0 0.51.5 1.02.0 1.52.5 2.03.0 2.5 0.5 0.01.0 0.51.5 1.02.0 1.52.5 2.03.0 2.5 0.0 3.0 0.5 0.0 1.4 95 1.4 β 102 11 11 10 10 9 1.41.3 γ 1.51.4 γ 9 9 8 1.61.5 8 1.6 single bid, T = 1 9.5 1.5 9 1.6 1.2 1.31.2 9.5 1.4 1.2 90 1.5 1.2 1.3 1.5 94 12 β 1.3 1.3 12 10 105 9.5 10 102 9.5 3.0 10 β 9 11 11.5 10 12 1.4 1.4 γ 90 1.51.4 1.5 β 9 98 1.3 γ 1.41.3 94 11 11.5 12 1.31.2 95 98 1.5 1.1 1.21.1 98 1.4 95 1.5 1.3 98 1.4 1.2 1.3 1.2 1.1 1.1 1.2 β 106 102 1.4 0 10 β 1.5 102 0 10 102 1.3 98 102 0 10 105 0 10 1.3 105 β 98 1.2 106 β 11.5 95 110 D1.6110 D1.6 11 β 11.5 β 11 B 1.5 B 1.5 105 0 0.0 multiple bids, T = 10 parameters of real auctions 1.6 A B 5.0 Evolutionary model for the auctions e = 20 4.0 A Moran-like model w (α) At generation e=0, agents take their strategies form 3.0 an arbitrary pdf (e) 1) they play the game. g (0) ( ) 2.0 2) a losing agent copies the strategy of the winner plus some random noise. e=5 1.0 3) e -> e + 1/N. Go back to point 1. e = 10 e=1 0.0 0 5.0 g (0) ( ) = const. 1.0 w (α) (e) 0.0 C 1.4 5 D e = 20 g (0) ( λ)= = const. 0.0 1.0 λ = 2.0 λ = 5.0 0.8 0 e=5 σ e = 10 10 -1 10 -2 10 -3 e=3 0.4 e=2 0.2 e=1 e=1 0.0 0 4 10 0.6 2.0 1.0 α 3 0.5 3.0 (e) w (α) BC 1.5 1.2 e = 20 4.0 2 random mutations 2.0 1.4 <α> A no mutations 1 1 2 w (e) e = 20 α 3 4 5 0.0 0 10 0 1 2101 e α 3 ( ) = prob. thatDthe winning strategy is 2.0 4 10 2 5 0.0 A Moran-like model 10 0 2.0 1.0 e=5 e = 10 σ (e) w (α) Evolutionary model for the auctions 0.5 3.0 λ = 0.0 λ = 2.0 λ = 5.0 10 -1 10 -2 10 -3 e=1 0.0 0 1.5 σ w (α) 10 (e) 10 0 -1 10 -2 10 -3 D e = 20 λ = 0.0 λ = 2.0 λ = 5.0 D e=3 <α> 0.0 1.2 1.0 2 10 2 random mutations 2.0 λ = 0.0 λ = 2.0 λ = 5.0 0.8 0.6 e=2 0.4 0.2 1 10 e 1 2 α 3 1.5 0.5 e 10 1.0 0.2 4102 5 2.0 1.0 1 0.0 1.2 0.6 0.0 0 10 0.5 0.8 0.4 0 0 1.0 e=1 10 10 1.5 1.0 0.0 5 1.4 1.0 0.5 4 no mutations 1.2 <α> α 3 <α> C 2.0 2 σ B 1 λ = 0.0 λ = 2.0 λ = 5.0 g (0) ( ) 0.0 0 10 10 e 1 Relative cost to winner A 2.5 Agents bid rationally, but ... C 2.5 Uniquebidhomes Uniquebidhomes 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0.0 3 10 Relative income of auctioneer Bidmadness B 6.0 10 4 10 5 10 6 10 7 5.0 4.0 4.0 3.0 3.0 2.0 2.0 1.0 1.0 0.0 3 10 10 4 10 5 10 6 10 10 D6.0 Uniquebidhomes 5.0 0.0 3 10 7 4 10 5 10 6 4 10 5 10 6 Bidmadness 0.0 3 10 10 Value of good r s, s <w s, s V V <Tc N +1 the economic return is always negative Do humans search as animals? 107 105 104 years ago 102 101 We are running some experiments... http://cgi.soic.indiana.edu/~i601levy/index.php Preliminary results Thanks L.A.N. Amaral, Northwestern University A. Baronchelli, City University London References Evolution of optimal Lévy-flight strategies in human mental searches ! F. Radicchi and A. Baronchelli ! Phys. Rev. E 85, 061121 (2012) Rationality, irrationality and escalating behavior in lowest unique bid auctions ! F. Radicchi, A. Baronchelli and L. A. N. Amaral ! PloS ONE 7, e29910 (2012) Lévy flights in human behavior and cognition! A. Baronchelli and F. Radicchi! Chaos Soliton. Fract. 56, 101-105 (2013)!