Peer-to-Peer Optimization in Large Unreliable Networks with Branch-and-Bound and Particle Swarms

Abstract

Abstract. Decentralized peer-to-peer (P2P) networks (lacking a GRID-style resource management and scheduling infrastructure) are an increasingly important computing platform. So far, little is known about the scaling and reliability of optimization algorithms in P2P environments. In this paper we present empirical results comparing two P2P algorithms for real-valued search spaces in large-scale and unreliable networks. Some interesting, and perhaps counter-intuitive findings are presented: for example, failures in the network can in fact significantly improve performance under some conditions. The two algorithms that are compared are a known distributed particle swarm optimization (PSO) algorithm and a novel P2P branch-and-bound (B&B) algorithm based on interval arithmetic. Although our B&B algorithm is not a black-box heuristic, the PSO algorithm is competitive in certain cases, in particular, in larger networks. Comparing two rather different paradigms for solving the same problem gives a better characterization of the limits and possibilities of optimization in P2P networks.

Evolving N-Body Simulations to Determine the Origin and Structure of the Milky Way Galaxy's Halo Using Volunteer Computing

“…In general these approaches are either sequential [11], [12], with distinct synchronization points; partially asynchronous, with few distinct synchronization points [13], [14]; or hybrid methods [15], [16], [17], [18], [19]. FGDO supports fully asynchronous versions of differential evolution, particle swarm optimization and genetic search, which differ from previous approaches as they remove all explicit synchronization points.…”

Validating Evolutionary Algorithms on Volunteer Computing Grids

“…Later, an inertia weight ω was added to the method by Shi and Eberhart to balance the local and global search capability of PSO [12] and is used in this work and by most modern PSO implementations. And recently, PSO has been shown to be effective in peer-to-peer computing environments by Bánhelyi et al [2]. The population of particles is updated iteratively as follows, where x is the position of the particle at iteration t, v is it's velocity, p is the individual best for that particle, and g is the global best position:…”

EvAg: a scalable peer-to-peer evolutionary algorithm

“…However, Genet Program Evolvable Mach (2010) 11:227-246 229 studying scalability in P2P optimization is not straightforward and is usually approached in two complementary ways, using real environments [25] or simulations [4]; either way presents its own advantages and drawbacks. On the one hand, performing a real massively distributed and decentralized experiment presents some challenges that, so far, pose a whole set of practical problems beyond the state of the art.…”

“…The drawback in this case is that simulations imply a certain number of assumptions about the real environment; hence, they have to be well stated (e.g., representing a pessimistic scenario as in [4]). In addition to approaches for real environments, some other works in the literature face the design of P2P EAs by means of simulations, focusing on the viability of the approaches rather than dealing with the harnessing of computing power.…”