Towards an Emergence-Driven Software Process for Agent-Based Simulation

Abstract

Abstract. In this paper we propose an emergence-driven software process for agent-based simulation that clarifies the traceability of micro and macro observations to micro and macro specifications in agent-based models. We use the concept of hyperstructures [1] to illustrate how micro and macro specifications interact in agent-based models, and we show that the reductionism/non-reductionism debate is important to understand the reliability of agent-based simulations. In particular, we show that the effort expended in the verification of agent-based simulations increases exponentially with the number of micro and macro specifications, and that the reliability assessment of non-anticipated results in simulation is in practice not possible. According to these results we claim to be impossible in practice to verify that an agent-based conceptual model has been implemented properly as a computational model, since one does not usually know what is the desired output a priori. We thus advocate that the classic process of verification, validation and exploration of non-anticipated results is not reliable in an agent-based simulation, and call into question the applicability of traditional software engineering methods to agentbased simulation.

Weak Interaction and Strong Interaction in Agent Based Simulations

“…Regarding this engineering process, a fundamental issue is raised using ABS: there is no consensus about the specifications that must be given to a CM (e.g. [4]). Thus, starting from a single CM and following this engineering process, several computational models (implementations) can be elaborated.…”

Validating Simulations

“…Virtually, the software notion of "functional requirement" does not exist in social simulation. If we are dealing with complex social models, it is virtually impossible to enumerate a priori an exhaustive list of requirements that a program should satisfy (David et al 2003). To the extent that the researcher does not know what to expect, program testing is not enough to show that the program properly reflects the intended model.…”

Seamless Integration and Testing for MAS Engineering

“…While the first four approaches (a-d) are considered formal or at least semi-formal, testing (e) is deemed to be an informal approach to verification. Further, MAS verification can be targeted at different levels, varying from fine-grained verification of agents at a logical level [3] to verification of emergent properties in a system [12]. Ferber [16] identifies three levels: (i) Agent level considers internal mechanisms and reasoning of an agent (ii) Group level consists in testing coordination mechanisms and interaction protocols of agents, and (iii) Society level checks for emergent properties or if certain rules and/or norms are complied within the society.…”