Catálogo de publicaciones - libros

Evolutionary algorithms based on “tags” can be adapted to induce cooperation in selfish environments such as peer-to-peer systems. In this approach, nodes periodically compare their utilities with random other peers and copy their behavior and links if they appear to have better utilities. Although such algorithms have been shown to posses many of the attractive emergent properties of previous tag models, they rely on the honest reporting of node utilities, behaviors and neighbors. But what if nodes do not follow the specified protocol and attempt to subvert it for their own selfish ends? We examine the robustness of a simple algorithm under two types of cheating behavior: a) when a node can lie and cheat in order to maximize its own utility and b) when a node acts nihilistically in an attempt to destroy cooperation in the network. For a test case representing an abstract cooperative application, we observe that in the first case, a certain percentage of such “greedy cheating liars” can actually improve certain performance measures, and in the second case, the network can maintain reasonable levels of cooperation even in the presence of a limited number of nihilist nodes.

Inspired by advances in evolutionary biology we extended existing evolutionary computation techniques and developed a self-organising, self-adaptable cellular system for multitask learning, called Evolvable Virtual Machine (EVM). The system comprises a specialised program architecture for referencing and addressing computational units (programs) and an infrastructure for executing those computational units within a global networked computing environment, such as Internet. Each program can be considered to be an agent and is capable of calling (co-operating with) other programs. In this system, complex relationships between agents may self-assemble in a symbiotic-like fashion. In this article we present an extension of previous work on the single threaded, single machine EVM architecture for use in global distributed environments. This paper presents a description of the extended Evolvable Virtual Machine (EVM) computational model, that can work in a global networked environment and provides the architecture for asynchronous massively parallel processing. The new computational environment is presented and followed with a discussion of experimental results.

This paper describes a car traffic control simulation realised in a decentralised way by message propagations: congested nodes (roads intersections) send speed-up or slow-down messages to neighbouring nodes. Different types of journeys have been modelled: regular car journeys, accidents and emergency cars journeys. These journeys have different lengths and speeds, and affect the system differently. Optimal values of parameters, used during the simulations for controlling the cars, have been determined through the use of a genetic algorithm (GA). This paper reports as well a preliminary experiment on different simulations realised with parameters values derived from the GA.