TY - GEN
T1 - On informational nudging and control of payoff-based learning
AU - Guers, Robin
AU - Langbort, Cedric
AU - Work, Dan
N1 - Funding Information:
★ This work was supported in part by an ISAE alumni grant and, in part, by the Office of the Vice–Chancellor for Research at UIUC, through the In3 Project “What Your Infrastructure Wants”. C.L. thanks California PATH at UC Berkeley for their hospitality while this work was initiated.
Copyright:
Copyright 2013 Elsevier B.V., All rights reserved.
PY - 2013
Y1 - 2013
N2 - We investigate a model of informational nudging in a context inspired by repeated games in traffic. Starting from a simple payoff-based learning model for an individual decision-maker (DM) choosing among multiple alternatives, we introduce a recommender who provides possibly misleading payoff information for unchosen options, so as to drive the DM's preferences to a desired equilibrium. This kind of white lie on the part of the recommender can be seen as an informational nudge in the sense of Thaler & Sunstein, and may thus arguably present some benefits over monetary incentive- based strategies for the purposes of planning. Considering the fluid limit of our simplified model, we show that the recommender can create (but not necessarily globally stabilize) any outcome he desires using constant lying strategies. We also identify a framing effect, in the sense that lies about the least favorable option has a different effect compared to lies on most favorable option.
AB - We investigate a model of informational nudging in a context inspired by repeated games in traffic. Starting from a simple payoff-based learning model for an individual decision-maker (DM) choosing among multiple alternatives, we introduce a recommender who provides possibly misleading payoff information for unchosen options, so as to drive the DM's preferences to a desired equilibrium. This kind of white lie on the part of the recommender can be seen as an informational nudge in the sense of Thaler & Sunstein, and may thus arguably present some benefits over monetary incentive- based strategies for the purposes of planning. Considering the fluid limit of our simplified model, we show that the recommender can create (but not necessarily globally stabilize) any outcome he desires using constant lying strategies. We also identify a framing effect, in the sense that lies about the least favorable option has a different effect compared to lies on most favorable option.
KW - Learning
KW - Stochastic approximation
KW - Traffic control
UR - http://www.scopus.com/inward/record.url?scp=84886629104&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84886629104&partnerID=8YFLogxK
U2 - 10.3182/20130925-2-DE-4044.00037
DO - 10.3182/20130925-2-DE-4044.00037
M3 - Conference contribution
AN - SCOPUS:84886629104
SN - 9783902823557
T3 - IFAC Proceedings Volumes (IFAC-PapersOnline)
SP - 69
EP - 74
BT - 4th IFAC Workshop on Distributed Estimation and Control in Networked Systems, NecSys 2013 - Proceedings
PB - IFAC Secretariat
T2 - 4th IFAC Workshop on Distributed Estimation and Control in Networked Systems, NecSys 2013
Y2 - 25 September 2013 through 26 September 2013
ER -