Home

Synopsis

Research

Papers

Adaptive, learning agents

• Work Unit 2 (WU2)

• Objective

• Approach

Important examples for adaptive agents concern negotiation, profiling, brokering, pricing, monitoring, scheduling, and marketing. Implicit learning is a significant aspect of the research. Application examples are the following: adaptive negotiation, dynamic pricing, bidding, and sales agents that implicitly assess the objectives and behaviour of other parties and create win-win situations; or (fast) profiling of users for derivation of (temporary) profiles, e.g. for first-time or anonymous customers or to avoid behavioural analysis of customers. The first example may also sustain chain inversion.

For the research, large computations and simulations on computer systems are executed, and a number of cases is modeled and investigated.

• Currently available deliverables

[0D2.2] Classification and Filtering with Trigrams and Evolutionary Nearest Neighbour Classifiers(restricted access only)
Scientific report on adaptive systems, concerning information-filtering and profiling based on evolutionary and heuristic methods. (2000 Q2)

[1D2.1] Scientific Techniques for Interactive Profiling. (restricted access only)
This survey paper gives an overview of adaptive information filtering and interactive profiling techniques both in scientific research and commercial applications. After clarifying the relation between information retrieval, information filtering, recommendation and user profiling the essential tasks in adaptive information filtering are outlined. When designing efficient profiling systems it is necessary to distinquish between two profiling types, (inter)active information maximizing approaches and unobtrusively observing ones. Scientific techniques for filtering and profiling are discussed in more detail, followed by a brief overview of research projects and existing filtering and recommending systems. Finally we sketch out prospective extensions of current commercial applications and further lines of research covering query learning and use of unlabeled data for efficient user profiling. (2001 Q2)

[1D2.2] Competitive Market-based Allocation of Consumer Attention Space: Concepts and Validation of Casy. (restricted access only)
The amount of attention space available for recommending suppliers to consumers on e-commerce sites is typically limited. We present a competitive distributed recommendation mechanism based on adaptive software agents for efficiently allocating the "consumer attention space", or banners. In the example of an electronic shopping mall, the task is delegated to the individual shops, each of which evaluates the information that is available about the consumer and his or her interests (e.g. keywords, product queries, and available parts of a profile). Shops make a monetary bid in an auction where a limited amount of "consumer attention space" for the arriving consumer is sold. Each shop is represented by a software agent that bids for each consumer. This allows shops to rapidly adapt their bidding strategy to focus on consumers interested in their offerings. For various basic and simple models for on-line consumers, shops, and profiles, we demonstrate the feasibility of our system by evolutionary simulations as in the field of agent-based computational economics (ACE). We also develop adaptive software agents that learn bidding-strategies, based on neural networks and strategy exploration heuristics. Furthermore, we address the commercial and technological advantages of this distributed market-based approach. The mechanisme we describe is not limited to the example of the electronic shopping mall, but can easily be extended to other domains. (2002 Q1-3)

[2D2.2] Bundling and Recommendation for Information Brokerage. (restricted access only)
In this paper, we discuss some of the consequences on-line dynamic bundling and/or pricing of (information) goods, and (automatic) recommender systems can have for information brokerage. We argue that dynamic bundling/pricing enhances especially the value extracting (or profit generating) capacity of an information broker. Recommerder systems, on the other hand, enhance through, for example, customer lock-in especially the value generating capacity of an information broker. More traditional (automatic) recommender systems have a number of drawbacks. We outline how recommendation based on sales statictics can circumvent these difficulties. We discuss especially the advantages and challenges of integrating dynamic bundling/pricing into such recommender systems. Keywords: value creation; value extraction; information brokerage; dynamic pricing; recommender systems.

[2D2.1] An Agent-Based Simulation for Market-Based Consumer Attention Allocation. (restricted access only)
In today's society consumers are exceedingly overwhelmed with both relevant and irrelevant information, the latter becoming more and more of a problem.This is especially pronounced on the Internet, where many advertisers attempt to reach potential customers. Lately, however, the traditional and undirected advertisements in the form of banners have shown to be less effective than predicted profit-wise. As a result, more and more companies focus on presenting targeted ads, which take into account information like the consumer's background and presumed product proferences. Displaying fewer but more relevant ads shown to be more effective, and as a result consumer-level marketinginformation has become a valueable asset. (2003 Q1-2)

[3D2.1] Bundling and Pricing for Information Brokerage: Customer Satisfaction as a Means to Profit Op timization. (restricted access only)
Traditionally, the study of on-line dynamic pricing and bundling strategies for information goods is motivated by the value-extracting or profitgenerating potential of these strategies. In this paper we discuss the relatively overlooked potential of these strategies to on-line learn more about customers' preferences. Based on this enhanced customer knowledge an information broker can -by tailoring the brokerage services more to the demand of the various customer groups- persuade customers to engage in repeated transactions (i.e. generate customer lock-in). To illustrate the duscussion, we show by means of a basic consumer model how, with the use of on-line dynamic bundling and pricing algorithms, customer lock-in can occur. The lock-in occurs because the algorithms can both find appropriate prices and (from the customers' perspective) the most interesting bundles. In the conducted computer experiments we use an advanced genetic algorithm with a niching method to learn the most interesting bundles efficiently and effectively.

[3D2.2] Bringing Haggling Back to the Market Place: How to Combine Bundling, Fairness and Win-Win. (restricted access only)
We present a novel system for selling single-day subscriptions on bundles of news items. Within the system, a news provider bargains with potential buyers using autonomous software agents that negotiate on the users' behalf. The advantage of the developed system is that it allows for highly flexible and customized subscriptions. We introduce the novel approach of decomposing bargaining strategies into a cooperative strategy. Non-cooperative strategies aim at obtaining the largest share of the pie, whereas cooperative strategies seek to maximize win-win opportunities. These win-win opportunities arise because bargaining involves multiple issues and trade-offs can be made. We furthermore present two cooperative strategies, the orthogonal and the orthogonal-DF strategy. Together they lead to win-win outcomes.