@Proceedings{IDM2016,
title = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
editor = {Tatiana V. Guy and Miroslav Kárný and David Rios-Insua and David H. Wolpert},
publisher = {PMLR},
series = {Proceedings of Machine Learning Research},
volume = 58
}
@InProceedings{pmlr-v58-guy17a,
title = {NIPS Workshop on Imperfect Decision Makers 2016: {P}reface},
author = {Guy, Tatiana V.},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {i--iii},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/guy17a/guy17a.pdf},
url = {https://proceedings.mlr.press/v58/guy17a.html},
abstract = {The workshop aims, organising and programme committees, invited talks and panel
discussions are introduced.}
}
@InProceedings{pmlr-v58-buckmann17a,
title = {Decision Heuristics for Comparison:{H}ow Good Are They?},
author = {Buckmann, Marcus and Şimşek, Özgür},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {1--11},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/buckmann17a/buckmann17a.pdf},
url = {https://proceedings.mlr.press/v58/buckmann17a.html},
abstract = {Simple decision heuristics are cognitive models of human and animal decision
making. They examine few pieces of information and combine the pieces in
simple ways, for example, by considering them sequentially or giving them
equal weight. They have been studied most extensively for the problem of
\textitcomparison, where the objective is to identify which of a given
number of alternatives has the highest value on a specified (unobserved)
criterion. We present the most comprehensive empirical evaluation of decision
heuristics to date on the comparison problem. In a diverse collection of 56
real-world data sets, we compared heuristics to powerful statistical learning
methods, including support vector machines and random forests. Heuristics
performed surprisingly well. On average, they were only a few percentage
points behind the best-performing algorithm. In many data sets, they yielded
the highest accuracy in all or parts of the learning curve.
The first part of the supplement describes implementation details of the
algorithms tested; the second part describes the 56 public data sets used in
the empirical analysis. }
}
@InProceedings{pmlr-v58-lichtenberg17a,
title = {Simple Regression Models},
author = {Lichtenberg, Jan M. and Şimşek, Özgür},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {13--25},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/lichtenberg17a/lichtenberg17a.pdf},
url = {https://proceedings.mlr.press/v58/lichtenberg17a.html},
abstract = {Developing theories of when and why simple predictive models
perform well is a key step in understanding decisions of cognitively
bounded humans and intelligent machines. We are interested in
how well simple models predict in regression. We list and review
existing simple regression models and define new ones. We identify
the lack of a large-scale empirical comparison of these models with
state-of-the-art regression models in a predictive regression context.
We report the results of such an empirical analysis on 60 real-world data
sets. Simple regression models such as equal-weights regression routinely
outperformed state-of-the-art regression models, especially on small training-set sizes.
There was no simple model that predicted well in all data sets, but in nearly
all data sets, there was at least one simple model that predicted well.
The supplementary material contains learning curves for individual data sets that have
not been presented in the main article. It also contains detailed descriptions and source
descriptions of all used data sets.
}
}
@InProceedings{pmlr-v58-karny17a,
title = {Implementable Prescriptive Decision Making},
author = {Kárný, Miroslav},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {19--30},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/karny17a/karny17a.pdf},
url = {https://proceedings.mlr.press/v58/karny17a.html},
abstract = {The need for inspecting (ir)rationality in decision making (DM)
— the observed discrepancy between real and prescriptive DMs —
stems from omnipresence of DM in individuals’ and society life.
Active approaches try to diminish this discrepancy either by changing
behaviour of participants (DM subjects) or modifying prescriptive
theories as done in this text. It provides a core of unified merging
methodology of probabilities serving for knowledge fusion and information
sharing exploited in cooperative DM. Specifically, it unifies merging
methodologies supporting a flat cooperation of interacting self-interested
DM participants. They act without a facilitator and they are unwilling
to spare a non-negligible deliberation effort on merging. They are supposed
to solve their DM tasks via the fully probabilistic design (FPD) of
decision strategies. This option is motivated by the fact that FPD
is axiomatically justified and extends standard Bayesian DM.
Merging is a supporting DM task and is also solved via FPD.
The proposed merging formulation tries to be as general as possible
without entering into technicalities of measure theory. The results generalise
and unify earlier work and open a pathway to systematic solutions of
specific, less general, problems.}
}
@InProceedings{pmlr-v58-drachal17a,
title = {Forecasting Spot Oil Price Using Google Probabilities},
author = {Drachal, Krzysztof},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {31--40},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/drachal17a/drachal17a.pdf},
url = {https://proceedings.mlr.press/v58/drachal17a.html},
abstract = {In this paper DMA (Dynamic Averaging Model) is expanded by adding certain probabilities
based on Google Trends. Such a method is applied to forecasting spot oil prices (WTI).
In particular it is checked whether a dynamic model including stock prices in
developed markets, stock prices in China, stock prices volatility, exchange rates,
global economic activity, interest rates, production, consumption, import and level of
inventories as independent variables might be improved by including a certain measure
of Google searches. Monthly data between 2004 and 2015 were analysed. It was found that such a
modification leads to slightly better forecast. However, the weight ascribed to Google
searches should be quite small. Except that it was found that even unmodified DMA produced
better forecast than that based on futures contracts or naive forecast.}
}
@InProceedings{pmlr-v58-seckarova17a,
title = {Performance of {K}ullback-{L}eibler Based Expert Opinion Pooling for Unlikely Events},
author = {Sečkárová, Vladimı́ra},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {41--50},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/seckarova17a/seckarova17a.pdf},
url = {https://proceedings.mlr.press/v58/seckarova17a.html},
abstract = {The aggregation of available information
is of great importance in many branches of economics, social sciences. Often,
we can only rely on experts’ opinions, i.e. probabilities assigned to
possible events. To deal with opinions in probabilistic form, we focus on the
Kullback-Leibler (KL) divergence based pools: linear, logarithmic and
KL-pool. Since occurrence of events is subject to random influences of the
real world, it is important to address events assigned lower probabilities
(unlikely events). This is done by choosing pooling with a higher entropy
than standard linear or logarithmic options, i.e. the KL-pool. We show how
well the mentioned pools perform on real data using absolute error,
KL-divergence and quadratic reward. In cases favoring events assigned higher
probabilities, the KL-pool performs similarly to the linear pool and
outperforms the logarithmic pool. When unlikely events occur, the KL-pool
outperforms both pools, which makes it a reasonable way of pooling.}
}
@InProceedings{pmlr-v58-guy17b,
title = {Experimental Performance of Deliberation-Aware Responder in Multi-Proposer Ultimatum Game},
author = {Guy, Tatiana V. and Ruman, Marko and Hůla, František and Kárný, Miroslav},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {51--60},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/guy17b/guy17b.pdf},
url = {https://proceedings.mlr.press/v58/guy17b.html},
abstract = {The ultimatum game serves for studying various aspects of decision making
(DM). Recently, its multi-proposer version has been modified to study the influence
of deliberation costs. An optimising policy of the responder, switching between several
proposers at non-negligible deliberation costs, was designed and successfully tested
in a simulated environment. The policy design was done within the framework of
Markov Decision Processes with rewards also allowing to model the responder’s feeling
for fairness. It relies on simple Markov models of proposers, which are recursively
learnt in a Bayesian way during the game course. This paper verifies, whether the
gained theoretically plausible policy, suits to real-life DM. It describes experiments
in which this policy was applied against human proposers. The results – with eleven
groups of three independently acting proposers – confirm the soundness of this policy.
It increases the responder’s economic profit due to switching between proposers, in spite
of the deliberation costs and the used approximate modelling of proposers. Methodologically,
it opens the possibility to learn systematically willingness of humans to spent their
deliberation resources on specific DM tasks.}
}
@InProceedings{pmlr-v58-ganzfried17a,
title = {Optimal Number of Choices in Rating Contexts},
author = {Ganzfried, Sam},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {61--74},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/ganzfried17a/ganzfried17a.pdf},
url = {https://proceedings.mlr.press/v58/ganzfried17a.html},
abstract = {In many settings people must give numerical scores to
entities from a small discrete set. For instance, rating physical
attractiveness from 1–5 on dating sites, or papers from 1–10
for conference reviewing. We study the problem of understanding
when using a different number of options is optimal. For concreteness
we assume the true underlying scores are integers from 1–100.
We consider the case when scores are uniform random and Gaussian.
While in theory for this setting it would be optimal to use all 100
options, in practice this is prohibitive, and it is preferable to
utilize a smaller number of options due to humans’ cognitive limitations.
Our results suggest that using a smaller number of options than is typical
could be optimal in certain situations. This would have many potential
applications, as settings requiring entities to be ranked by humans are ubiquitous.}
}
@InProceedings{pmlr-v58-simsek17a,
title = {On Learning Decision Heuristics},
author = {Şimşek, Özgür and Buckmann, Marcus},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {75--85},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/simsek17a/simsek17a.pdf},
url = {https://proceedings.mlr.press/v58/simsek17a.html},
abstract = {Decision heuristics are simple models of human and animal decision making
that use few pieces of information and combine the pieces in simple ways, for example,
by giving them equal weight or by considering them sequentially. We examine how decision
heuristics can be learned—and modified—as additional training examples become available.
In particular, we examine how additional training examples change the variance in parameter
estimates of the heuristic. Our analysis suggests new decision heuristics, including
a family of heuristics that generalizes two well-known families: lexicographic heuristics
and tallying. We evaluate the empirical performance of these heuristics
in a large, diverse collection of data sets.
The supplementary material provides details on the random
forest implementation and describes the 56 public data sets used in the empirical analysis.
}
}
@InProceedings{pmlr-v58-benavoli17a,
title = {Quantum Rational Preferences and Desirability},
author = {Benavoli, Alessio and Facchini, Alessandro and Zaffalon, Marco},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {87--96},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/benavoli17a/benavoli17a.pdf},
url = {https://proceedings.mlr.press/v58/benavoli17a.html},
abstract = {We develop a theory of quantum rational decision making in the tradition of
Anscombe and Aumann’s axiomatisation of preferences on horse lotteries.
It is essentially the Bayesian decision theory generalised to
the space of Hermitian matrices. Among other things, this leads us to give
a representation theorem showing that quantum complete rational preferences
are obtained by means of expected utility considerations.}
}
@InProceedings{pmlr-v58-d-agostino17a,
title = {Rational Beliefs Real Agents Can Have – {A} Logical Point of View},
author = {D’Agostino, Marcello and Flaminio, Tommaso and Hosni, Hykel},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {97--109},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/d-agostino17a/d-agostino17a.pdf},
url = {https://proceedings.mlr.press/v58/d-agostino17a.html},
abstract = {The purpose of this note is to outline a framework for uncertain
reasoning which drops unrealistic assumptions about the agents’
inferential capabilities. To do so, we envisage a pivotal role for
the recent research programme of \emphdepth-bounded Boolean logics.
We suggest that this can be fruitfully extended to the representation
of rational belief under uncertainty. By doing this we lay the
foundations for a prescriptive account of rational belief, namely one
that \emphrealistic agents, as opposed to idealised ones, can feasibly act upon.}
}
@InProceedings{pmlr-v58-mahdavi17a,
title = {Hindsight Bias Impedes Learning},
author = {Mahdavi, Shaudi and Rahimian, M. Amin},
booktitle = {Proceedings of the NIPS 2016 Workshop on Imperfect Decision Makers},
pages = {111--127},
year = {2017},
editor = {Guy, Tatiana V. and Kárný, Miroslav and Rios-Insua, David and Wolpert, David H.},
volume = {58},
series = {Proceedings of Machine Learning Research},
month = {09 Dec},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v58/mahdavi17a/mahdavi17a.pdf},
url = {https://proceedings.mlr.press/v58/mahdavi17a.html},
abstract = {We propose a model that addresses an open question
in the cognitive science literature: How can we rigorously model
the cognitive bias known as hindsight bias such that
we fully account for critical experimental results?
Though hindsight bias has been studied extensively,
prior work has failed to produce a consensus theoretical model
sufficiently general to account for several key experimental results,
or to fully demonstrate how hindsight impedes our ability to learn
the truth in a repeated decision or social network setting. We
present a model in which agents aim to learn the quality of
their signals through repeated interactions with their environment.
Our results indicate that agents who are subject to hindsight
bias will always believe themselves to be high-type “expert” regardless
of whether they are actually high- or low-type.}
}