Building a Computational Digital Economy through Interdisciplinary Research

by Petros Kavassalis and Konstantin Popov

Researchers from the Swedish Institute of Computer Science (SICS) and the ATLANTIS Group at the University of Crete are forming a multidisciplinary team to study the complex issues of consumer decision-making in the online world. Their approach combines modern computer modelling and simulation methods with techniques from the emerging field of neuroeconomics.

Consumers on the Web take advantage of the fact that information is available in excess. They search and gather product information, compare prices and check product availability. They access free reviews in online consumer forums and share product experience in slashdot-like agoras. Several economists are now talking about a new marketing world in which consumers enjoy a superior information endowment, and suggest that businesses should adopt high-differentiation strategies as a way of reducing the importance of price in consumer choice.

This 'information excess' introduces a significant effect: the process of information perception and evaluation might determine, almost 'bias', the entire consumer decision-making process. Consumers on the Web will increasingly have a quasi-realistic experience with products, and may make their decision over the Web alone. In addition, various self-reinforcing, information-propagation mechanisms, such as the Online Feedback Mechanisms, enable a complex choice process and act as positive feedbacks into consumer choice behaviour.

In studying such a complex system, there are many questions to be answered. How are individual preferences being constructed and developed? How do these preferences become coherent in such an information-rich environment? Does the 'information excess' increase or decrease the uncertainty with respect to the value one should give to various product attributes? What is the impact of previous consumer experience, if any, on preference stability? How efficient can online feedback mechanisms be in making consumers organizationally effective and creative under these conditions? The challenge is to develop behavioural models that can help to understand how consumers make choices in information-rich environments, and to investigate efficient cyber-consumer welfare policies and effective business strategies in online markets.

Modelling Digital Economy
Algorithmic game theory can provide mechanisms for understanding how online agents with diverse objectives can coordinate adaptively and arrive at a stable equilibrium. Alternatively, complex systems methods allow for a greater variety of individual behaviour and for simplicity with respect to the reasoning attributed to the individuals. In addition, they allow for individual preferences and interaction links between individuals to be embedded in networks and distributed across the population with non-uniform probability.

Computer-based modelling and simulation can be used to study online markets as 'non-cooperative games'. Neuroeco-nomics can help to investigate the physical mechanisms by which the human neuroarchitecture accomplishes product selection in the online world. These two areas together can yield very realistic behavioural models of 'intelligent' agents representing cyber-consumers. The structure of these agents will be more than a framework for implementing rational or steady-state adaptive behaviour, and will reflect the organization of the mind when consumers make decisions. Hence, it is expected to: (i) possess capabilities that single out particular product attributes as meaningful, (ii) take into account advice from other consumers, and (iii) feature proactiveness and deploy internal commitment ability. Agents' structure should therefore include a number of cognitive elements (sensation, cognitive control, learning, memory, processes for social behaviour etc) and a cognitive architecture, such as ACT-R (see Links) that coherently integrates components. Decision-making will arise from the interaction of these cognitive processes, as constrained by the architecture.

Intelligent Agents
'Intelligent' cyber-consumer agents reason about products which, in turn, are represented as vectors of different functional characteristics. This allows for product differentiation extending far beyond typical spatial differentiation, with uniformly distributed consumer preferences and products that are spatially differentiated along a simple unit circle. Finally, we model the strategies of online marketers to be associated with products and exerted to influence the competition game.

Simulation of these models will generate highly dynamic, non-linear, experimental online markets with a high density of circular dependencies and information flows, the study of which can enrich our understanding of real Web markets. Data from the real Web, such as product listings with full descriptive characteristics, and product rankings taken from online forums, should be integrated into the system during simulation, making the world in which our artificial agents live more realistic. Pursuing this line of experimentation, the simulation of a digital economy is intended to effectively reproduce several structural properties of the real Web, and unfold drivers for the Web's deployment and growth to identify causal (economic) forces at work that affect its design.

Integrating Data into Simulation
Two types of data can enhance simulation. The first is information on the possible behaviour of cyber-consumer agents, collected through neuro-experiments with volunteer participants acting in quasi-realistic conditions. The second is data collected from the real Web that will be used to populate the local environment of the individual agents. This environment is, of course, altered by the agents' actions during the course of the simulation. We propose to use a simple(r) model of the system as the context for analysis and exploitation of data. Whenever state and inputs of a simulated agent match the available data, its outputs should also match the data. Analysis can include inferring relationships between elements of data, validation of hypothesized models of individual agents produced by model designers, and maybe even automated generation and calibration of models using machine-learning techniques. In this context, analysis of data and validation of models of agents become integrated into the agent-based simulation. Learning of agent behaviour through simulation has already been proposed by 'participatory simulation', where humans substitute for model agents during simulation runs. However, participatory simulation deals with humans that entirely substitute for model agents, whereas in our case, both experimental and publicly available data are analysed and continuously used during 'pure' agent-based simulation.

A general-purpose concurrent programming system like Mozart, can be used as a platform for developing simulations with cognitive agents and integration of real-world data. Model agents are conceptually concurrent. Their concurrent implementation can be exploited for distributed/parallel execution aiming at scalability and performance, which becomes crucial with large-scale, computationally intensive models. Integration of cognitive architectures and data analysis into model agents requires flexibility of a general-purpose programming language.


Please contact:
Petros Kavassalis
University of Crete, Greece

Konstantin Popov
SICS, Sweden