CALL FOR CONTRIBUTIONS

ICML Workshop on Markets, Mechanisms, and Multi-Agent Models: Examining the Interaction of Machine Learning and Economics

http://icml2012marketswkshop.pbworks.com/

Important Dates:

----------------

Deadline for submissions: April 20, 2012  May 4, 2012 (DEADLINE EXTENDED)

Notification of Acceptance: May 18, 2012

Submissions 
----------- 
We welcome submissions in the areas of multi-agent systems, aggregation mechanisms and crowdsourcing, belief elicitation, reward based systems, game theory and mechanism design, online trading and portfolio selection, prediction markets, and auctions or other areas of relevance. In each of these areas there are often transactional or reward mechanisms that can be thought of in economic terms.

Overview:

---------

Many of society’s greatest accomplishments are in large part due to the facility of markets. Markets and other allocation mechanisms have become necessary tools of the modern age, and they have been key to facilitating the development of complex structures, advanced engineering, and a range of other improvements to our collective capabilities. Much work in economics has been done to demonstrate that markets can, in aggregate, function very well even when the individual participants are noisy, irrational or myopic.

In terms of aims and benefits, the design of machine learning techniques has much in common with the development of market mechanisms: information aggregation, maximal efficiency, scalability, and, more recently, decentralization. Current machine learning algorithms are often single goal methods, built from simple homogeneous units by one person or individual groups. Perhaps looking to the organisations of economies may help in moving beyond the current centralised design of most machine learning methods. Allowing agents with different opinions, approaches or methods to enter and leave the market, to interact, and to adapt to changes can have many benefits. For example it may enable us to develop methods that provide continuous improvement on complex problems, reuse results by improving on previous outcomes rather than building bigger models from scratch, and adapting to changes.

There are many relationships between machine learning methods, Bayesian decision theory, risk minimisation, economics, statistical physics and information theory that have been known for some time. There are also many open questions regarding the full nature and impact of these connections. This workshop will explore these connections from many different directions. Some example topics are given below.

Various Topics:

-------

1) Prediction markets as a tool for learning and aggregation. 

The problem of how to aggregate information across various sources and types of data, as well as aggregating the predictions of different hypotheses or classifiers, has been of central focus in the learning community. But much of this work has assumed that the process can be centralized, that all of the inputs are available to a single algorithm. Prediction markets have been proposed as a tool for aggregating information distributed across various agents, each holding their own incentives and private beliefs. There is already a budding literature connecting these two ideas, and we would like to explore this further.

2) Learning in problems of mechanism design. 

Economists have studied the problem of designing of auctions (mechanisms) for decades, and in the past ten years there is an increasing amount of work being done within the CS community as well. The key problem is how to design the allocation of a set of goods, based on the inputs from a number of agents with private valuations of the goods, in order that the mechanism is incentive compatible, i.e., such that the agents have no reason to lie about their interests. As many authors have observed, this can be seen as a problem of prediction, since our goal is to answer in some optimal way, “what shall I charge for these goods?” There are already several papers that have made connections along these lines.

3) Prediction and learning in ad auctions. 

The internet advertising market is a perfect example where we have a learning problem, “predicting the likelihood that a user with given features will click on a given ad,” mixed with an efficiency problems, “determining how to allocate ad inventory to page views across different venues.” Making the problem harder, these decisions often need to be made online, in a distributed fashion, and in less than 100ms.

4) Online trading and portfolio selection

A number of techniques used in finance -- Markowitz portfolio selection, adaptive hedging strategies for pricing options and other derivatives, so-called universal portfolio algorithms -- are essentially online learning algorithms. There has been some work to bring out the connections between ML and finance, but we believe this is a very open field.

5) Relating Markets and Machine Learning Methods. 

There is a direct equivalence between expected utility theory in economics, and decision theory as used in machine learning. Entropy, in information theoretic and statistical physics senses is directly relevant in stock market models: “the growth rate of wealth is a dual of the entropy rate of a stock market” (Cover & Thomas 2nd. Ed. 2006). We welcome contributions that explore the striking relationships between these fields. 

6) Transactional Communication in Multi-agent Systems. 

In multi agent systems, such as those related to sensor networks, where there is a shared goal, there is a system of distributed information and action. In order to perform well, agents must communicate information, and take actions, both of which are costly. With autonomous agents the communication of information and action is transactional, and can be considered using market, game-theoretic or decision process frameworks.  We are interested in approaches that explore the connections between multi-agent and economic systems. 

7) Other -- feel free to email the organizers regarding additional topics.