Shopb ots and Priceb ots: Ho w will b ots aect m a rk - - PDF document
▶
SMART_READER_LITE
LIVE PREVIEW
Shopb ots and Priceb ots: Ho w will b ots aect m a rk - - PDF document
Jan 20, 2024 147 likes •324 views
Shopb ots and Priceb ots: Ho w will b ots aect m a rk ets? Am y Greenw ald and Je Kepha rt IBM Institute fo r Advanced Com m erce IBM Thom as J. W atson Resea rch Center Ha wtho rne, New Y o rk
SLIDE 1Shopb
ts
and Priceb
ts:
Ho w will b
ts
aect m a rk ets? Am y Greenw ald and Je Kepha rt IBM Institute fo r Advanced Com m erce IBM Thom as J. W atson Resea rch Center Ha wtho rne, New Y
rk
August 5, 1999
SLIDE 2Econom ics
f
Info rm a tion Geo rge Stigler [1961]
p
rice disp ersion is attributed to costly sea rch p ro cedures Shopb
ts
T
da
y (Y esterda y!)
shopb
ts
sp ecialize in collecting and distributing p rice info rm a tion at lo w cost Priceb
ts
T
m
rro
w (T
da
y!)
autom
ated agents that set p rices in attem pt to m axim ize p rots fo r sellers, just as shopb
ts
seek to m inim ize costs fo r buy ers 1
SLIDE 3Overview Sellers
Gam
e-Theo retic Equilib rium
Strategic
Priceb
t
Dynam ics Buy ers
Gam
e-Theo retic Equilib rium
Rational
Buy er Dynam ics 2
SLIDE 4Gam e-Theo retic Priceb
t
Strategy Mixed strategy Nash equilib rium Rational p riceb
ts
cho
se
p rices at random acco rding to p robabilit y distribution. 5 p riceb
ts,
w 1 = 0:2, w 2 + w 5 = 0:8.
0.5 0.6 0.7 0.8 0.9 1.0 5 10 15 20 p f(p)
w2=0.8 w2=0.4 w2=0.0
0.5 0.6 0.7 0.8 0.9 1.0 5 10 15 20
20 p riceb
ts,
w 1 = 0:2, w 2 + w 20 = 0:8.
0.5 0.6 0.7 0.8 0.9 1.0 5 10 15 20 p f(p)
w2=0.8 w2=0.4 w2=0.0
Do adaptive (not necessa rily rational) p riceb
ts
lea rn gam e- theo retic equilib rium
f
stage gam e
ver
rep eated pla ys? 3
SLIDE 5Info rm ed, Adaptive Priceb
t
Strategy My
pically-optim
al (MY) Strategy Rational b est-resp
nse
to
thers'
current p rices, given buy er dem a nd function 5 MY Priceb
ts
. . .
1 2 3 4 5
×106
0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 Time Price
Valuation Production Cost
5 MY
w1=0.25, w2=0.50, w5=0.25
1 2 3 4 5
×106
0.00 0.05 0.10 0.15 0.20 0.25 0.30 Time Profit
Average MY Profit = 0.0524
MY
p rots (0.0524) m
r
e than t wice GT p rots (0.025);
but
instabilities in the fo rm
f
cyclical p rice w a rs a rise;
and
MY p riceb
t
requires kno wledge
f
buy er dem and and
ther
sellers' p rices, which m a y b e costly to
btain.
4
SLIDE 6Naive, Adaptive Priceb
t
Strategy Derivative-follo wing (DF) Strategy Adjust p rice in sam e direction as long as p rot increases;
therwise
reverse the direction
f
p rice adjustm ent. 5 DF Priceb
ts
. . .
1 2 3 4 5 6 7 8 9 10
×106
0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 Time Price
Valuation Production Cost
5 DF
w1=0.25, w2=0.50, w5=0.25
1 2 3 4 5 6 7 8 9 10
×106
0.00 0.05 0.10 0.15 0.20 0.25 0.30 Time Profit
Average DF Profit = 0.0733
T
acit collusion results: i.e., an eective ca rtel despite no actual com m unication!
Average
p rot is nea rly 3 tim es that
f
GT p riceb
ts.
P erfect ca rtel w
uld
achieve p rot
f
0.1 p er p riceb
t.
Requires
no kno wledge
f
sellers' p rices
r
buy er dem and; p rice-setting m echanism based
n
histo rical
bservations.
5
SLIDE 7Info rm ed vs. Naive Priceb
ts
Intro duce 1 MY p riceb
t
into group
f
4 DF p riceb
ts
. . .
1 2 3 4 5 6 7 8 9 10
×106
0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 Time Price
Valuation Production Cost
1 MY vs. 4 DF
w1=0.25, w2=0.50, w5=0.25
1 2 3 4 5 6 7 8 9 10
×106
0.00 0.05 0.10 0.15 0.20 0.25 0.30 Time Profit
Average MY Profit = 0.1209 Average DF Profit = 0.0523
. . . and it will exploit them m ercilessly , stealing their p rots, ea rning m
re
than t wice (0.121) what they do (0.052)! 6
SLIDE 8Q-Lea rning Priceb
ts
W atkins, 1989 Reinfo rcement Lea rning Schem e 2 MY Priceb
ts
. . . 2 Q Priceb
ts
. . .
1 2 3 4 5 6 7 8 9 10
×106
0.0 0.1 0.2 0.3 0.4 0.5 0.6 Time Price
Mean Valuation Production Cost
1 MY vs. 1 MY
w1=0.25, wS=0.75
1 2 3 4 5 6 7 8 9 10
×106
0.0 0.1 0.2 0.3 0.4 0.5 0.6 Time Price
Mean Valuation Production Cost
1 Q vs. 1 Q
w1=0.25, wS=0.75
Q
p riceb
ts
detect and abandon p rice w a rs ea rly
n
Q
p rots (0.125, 0.117) exceed MY p rots (0.089, 0.089) 7
SLIDE 9No External Regret Priceb
ts
F reund and Schapire, 1995 Probabilistic Up dating Scheme 2 NER Priceb
ts
. . .
1 2 3 4 5 6 7 8 9 10
×104
0.5 0.6 0.7 0.8 0.9 1.0 No External Regret Learning Time Price 2 NER Sellers, wS = 0.75 2 4 6 8 10
×104
0.5 0.6 0.7 0.8 0.9 1.0 No External Regret Responsive Learning Time Price 2 NER Sellers, wS = 0.75
NER
p riceb
ts
cycle through p rices exp
nentially
resp
nsive
NER p riceb
ts
engage in lim ited p rice w a rs 8
SLIDE 10No Internal Regret Priceb
ts
F
ster
and V
hra,
1997 Converge to Co rrelated Equilib rium 2 NIR Priceb
ts
. . .
0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.2 0.4 0.6 0.8 1.0
wB = 0.9 wB = 0.75 wB = 0.5 wB = 0.25 wB = 0.1
No Internal Regret Learning Price Probability 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.2 0.4 0.6 0.8 1.0
wB = 0.9 wB = 0.75 wB = 0.5 wB = 0.25 wB = 0.1
Nash Equilibrium Price Probability
. . . lea rn Nash equilib rium! 9
SLIDE 11Rational Buy er Strategy T
tal
Buy er Exp enditure = Exp ected Price + Sea rch Costs Buy er Price Distributions
xed and adaptive buy er b ehavio r can lead to strategies
ther
than just sea rch-2 co-existing with sea rch-1. 14
SLIDE 16F uture W
rk
Study
dynam ics
f
adaptive buy ers and sellers
Investigate
strategic interpla y
f
shopb
t
p ricing
Dynam
ic p ricing
f
p rice and p ro duct info rm ation in full- edged economy
f
soft w a re agents, consisting
f
buy ers, sellers, and econom ically m
tivated
shopb
ts
Shopb
t
Econom ics fo rm s pa rt
f
the Info rm ation Econom ies p roject at IBM Resea rch Institute fo r Advanced Com m erce. The p roject goal is to: accurately describ e and p redict collective interactions
f
billions
f
economically m
tivated
soft w a re agents, and use insights so gained to design agent strategies, p roto cols, and infrastructures. Project description and resea rch pap ers available at: www.research .ib m.c
