On coevolutionary genetic algorithms

 The use of evolutionary computing techniques in coevolutionary/multi-agent systems is becoming increasingly popular. This paper presents simple models of the genetic algorithm in such systems, with the aim of examining the effects of different types of interdependence between individuals. Using the model it is shown that, for a fixed amount of interdependence between coevolving individuals, the existence of partner gene variance and the level at which fitness is applied can have significant effects, as does the evaluation partnering strategy used.     

Simple models of coevolutionary genetic algorithms

The use of evolutionary computing techniques in coevolutionary/multiagent systems is becoming increasingly popular. This paper presents some simple models of the genetic algorithm in such systems, with the aim of examining the effects of different types of interdependence between individuals. Using the models, it is shown that for a fixed amount of interdependence between homogeneous coevolving individuals, the existence of partner gene variance, gene symmetry, and the level at which fitness is applied can have significant effects. Similarly, for heterogeneous coevolving systems with fixed interdependence, partner gene variance and fitness application are also found to have a significant effect, as is the partnering strategy used.     

Local convergence analysis of FastICA and related algorithms

The FastICA algorithm is one of the most prominent methods to solve the problem of linear independent component analysis (ICA). Although there have been several attempts to prove local convergence properties of FastICA, rigorous analysis is still missing in the community. The major difficulty of analysis is because of the well-known sign-flipping phenomenon of FastICA, which causes the discontinuity of the corresponding FastICA map on the unit sphere. In this paper, by using the concept of principal fiber bundles, FastICA is proven to be locally quadratically convergent to a correct separation. Higher order local convergence properties of FastICA are also investigated in the framework of a scalar shift strategy. Moreover, as a parallelized version of FastICA, the so-called QR FastICA algorithm, which employs the QR decomposition (Gram-Schmidt orthonormalization process) instead of the polar decomposition, is shown to share similar local convergence properties with the original FastICA.     

Performance analysis of distributed deadlock detection algorithms

The paper presents a probabilistic performance analysis of a deadlock detection algorithm in distributed systems. Although there has been extensive study on deadlock detection algorithms in distributed systems, little attention has been paid to the study of the performance of these algorithms. Most work on performance study has been achieved through simulation but not through an analytic model. Min (1990), to the best of our knowledge, made the sole attempt to evaluate the performance of distributed deadlock detection algorithms analytically. Being different from Min's, our analytic approach takes the time-dependent behavior of each process into consideration rather than simply taking the mean-value estimation. Furthermore, the relation among the times when deadlocked processes become blocked is studied, which enhances the accuracy of the analysis. We measure performance metrics such as duration of deadlock, the number of algorithm invocations, and the mean waiting time of a blocked process. It is shown that the analytic estimates are nearly consistent with simulation results     

A convergence analysis of generalized hill climbing algorithms

Generalized hill climbing (GHC) algorithms provide a well-defined framework for describing the performance of local search algorithms for discrete optimization problems. Necessary and sufficient convergence conditions for GHC algorithms are presented. These convergence conditions are derived using a new iteration classification scheme for GHC algorithms. The implications of the necessary and the sufficient convergence conditions fur GHC algorithms with respect to existing convergence theory for simulated annealing are also discussed     

Sufficient conditions for the convergence of a class of nonlinear distributed consensus algorithms

This paper is concerned with the convergence of a class of continuous-time nonlinear consensus algorithms for single integrator agents. In the consensus algorithms studied here, the control input of each agent is assumed to be a state-dependent combination of the relative positions of its neighbors in the information flow graph. Using a novel approach based on the smallest order of the nonzero derivative, it is shown that under some mild conditions the convex hull of the agents has a contracting property. A set-valued LaSalle-like approach is subsequently employed to show the convergence of the agents to a common point. The results are shown to be more general than the ones reported in the literature in some cases. An illustrative example demonstrates how the proposed convergence conditions can be verified.     

A distributed Cooperative coevolutionary algorithm for multiobjective optimization

Recent advances in evolutionary algorithms show that coevolutionary architectures are effective ways to broaden the use of traditional evolutionary algorithms. This paper presents a cooperative coevolutionary algorithm (CCEA) for multiobjective optimization, which applies the divide-and-conquer approach to decompose decision vectors into smaller components and evolves multiple solutions in the form of cooperative subpopulations. Incorporated with various features like archiving, dynamic sharing, and extending operator, the CCEA is capable of maintaining archive diversity in the evolution and distributing the solutions uniformly along the Pareto front. Exploiting the inherent parallelism of cooperative coevolution, the CCEA can be formulated into a distributed cooperative coevolutionary algorithm (DCCEA) suitable for concurrent processing that allows inter-communication of subpopulations residing in networked computers, and hence expedites the computational speed by sharing the workload among multiple computers. Simulation results show that the CCEA is competitive in finding the tradeoff solutions, and the DCCEA can effectively reduce the simulation runtime without sacrificing the performance of CCEA as the number of peers is increased.     

Fitness function shaping in multiagent cooperative coevolutionary algorithms

Coevolution is a promising approach to evolve teams of agents which must cooperate to achieve some system objective. However, in many coevolutionary approaches, credit assignment is often subjective and context dependent, as the fitness of an individual agent strongly depends on the actions of the agents with which it collaborates. In order to alleviate this problem, we introduce a cooperative coevolutionary algorithm which biases the evolutionary search as well as shapes agent fitness functions to promote behavior that benefits the system-level performance. More specifically, we bias the search using a hall of fame approximation of optimal collaborators, and shape the agent fitness using the difference evaluation function. Our results show that shaping agent fitness with the difference evaluation improves system performance by up to 50 %, and adding an additional fitness bias improves performance by up to 75 % in our experiments. Finally, an analysis of system performance as a function of computational cost demonstrates that this algorithm makes extremely efficient use of computational resources, having a higher performance as a function of computational cost than any other algorithm tested.     

Network-based distributed planning using coevolutionary agents: architecture and evaluation

A novel evolutionary planning framework (coevolutionary virtual design environment) particularly suited to distributed network-enabled design and manufacturing organizations is presented. The approach utilizes distributed evolutionary agents and mobile agents as principal object-oriented software entities that support a network-efficient evolutionary exploration of planning alternatives in which successive populations systematically select planning alternatives that reduce cost and increase throughput. This paper presents the architecture of the coevolutionary virtual design environment, and examines the network-based performance of the coevolutionary algorithms that execute in this environment. Simulation analysis examines the percentage convergence error and percentage computational advantage comparing the distributed network-based implementation to a centralized network-based implementation. The algorithms and architectures are evaluated in a realistic network setting and analyzed using models of network delays and processing times.     

Comparison between two coevolutionary feature weighting algorithms in clustering

Feature weighting is an aspect of increasing importance in clustering because data are becoming more and more complex nowadays. In this paper, we propose two new feature weighting methods based on coevolutive algorithms. The first one is inspired by the Lamarck theory (inheritance of acquired characteristics) and uses the distance-based cost function defined in the LKM algorithm as fitness function. The second method uses a fitness function based on a new partitioning quality measure. It does not need a distance-based measure. We compare classical hill-climbing optimization with these new genetic algorithms on three data sets from UCI. Results show that the proposed methods are better than the hill-climbing based algorithms. We also present a process of hyperspectral remotely sensed image classification. The experiments, corroborated by geographers, highlight the benefits of using coevolutionary feature weighting methods to improve knowledge discovery process.     

Virtual player design using self-learning via competitive coevolutionary algorithms

The Google Artificial Intelligence (AI) Challenge is an international contest the objective of which is to program the AI in a two-player real time strategy (RTS) game. This AI is an autonomous computer program that governs the actions that one of the two players executes during the game according to the state of play. The entries are evaluated via a competition mechanism consisting of two-player rounds where each entry is tested against others. This paper describes the use of competitive coevolutionary (CC) algorithms for the automatic generation of winning game strategies in Planet Wars, the RTS game associated with the 2010 contest. Three different versions of a prime algorithm have been tested. Their common nexus is not only the use of a Hall-of-Fame (HoF) to keep note of the winners of past coevolutions but also the employment of an archive of experienced players, termed the hall-of-celebrities (HoC), that puts pressure on the optimization process and guides the search to increase the strength of the solutions; their differences come from the periodical updating of the HoF on the basis of quality and diversity metrics. The goal is to optimize the AI by means of a self-learning process guided by coevolutionary search and competitive evaluation. An empirical study on the performance of a number of variants of the proposed algorithms is described and a statistical analysis of the results is conducted. In addition to the attainment of competitive bots we also conclude that the incorporation of the HoC inside the primary algorithm helps to reduce the effects of cycling caused by the use of HoF in CC algorithms.     

On the convergence of the holistic analysis for EDF distributed systems

Dynamic scheduling techniques, and EDF (Earliest Deadline First) in particular, have demonstrated their ability to increase the schedulability of real time systems compared to fixed-priority scheduling. In distributed systems, the scheduling policies of the processing nodes tend to be the same as in stand-alone systems and, although few EDF networks exist, it is foreseen that dynamic scheduling will gradually develop into real-time networks. There are some response time analysis techniques for EDF scheduled distributed systems, mostly derived from the holistic analysis developed by Spuri. A major factor influencing the response time is the release jitter of each task, which is the maximum variation suffered by the release time of the task jobs. The convergence of the holistic analysis in the context of EDF distributed systems with shared resources had not been studied until now. There is a circular dependency between the task release jitter values, response times and the preemption level ceilings of shared resources. In this paper we present an extension of Spuri’s algorithm and we demonstrate that its iterative formulas are non-decreasing, even in the presence of shared resources. This result enables us to assert that the new algorithm converges towards a solution for the response times of the tasks and messages in a distributed system.¹     

A coevolutionary approach to adapt the genotype-phenotype map in genetic algorithms

This article introduces a coevolutionary approach to genetic algorithms (GAs) for exploring not only within a part of the solution space defined by the genotype-pheno-type map, but also the map itself. In canonical GAs with a fixed map, how large an area of the solution space can be covered by possible genomes, and consequently how better solutions can be found by a GA, rely on how well the geotype-phenotype map in designed, but it is difficult for designers of the algorithms to design the map without a priori knowledge of the solution space. In the proposed algorithm, the genotype-phenotype map is improved adaptively during the search process for solution candidates. It is applied to 3-bit deceptive problems such as of typical combinatorial optimazation problems. These are well known because their difficulty for GAs can be controlled by the genotype-phenotype map, and this shows a fairly good performance compared with a conventional GA. This work was presented in part at the Sixth International Symposium on Artificial Life and Robotics, Tokyo, January 15–17, 2001.     

Exponential convergence of adaptive identification and control algorithms

Output and equation error adaptive identification algorithms are shown to be exponentially convergent under a deterministic or stochastic persistently exciting (or spanning) condition on the system inputs together with several other standard conditions. An adaptive control algorithm is shown to be exponentially convergent under a deterministic or stochastic persistently exciting condition on the reference trajectory together with some standard conditions.     

Gradual distributed real-coded genetic algorithms

A major problem in the use of genetic algorithms is premature convergence. One approach for dealing with this problem is the distributed genetic algorithm model. Its basic idea is to keep, in parallel, several subpopulations that are processed by genetic algorithms, with each one being independent of the others. Making distinctions between the subpopulations by applying genetic algorithms with different configurations, we obtain the so-railed heterogeneous distributed genetic algorithms. These algorithms represent a promising way for introducing a correct exploration/exploitation balance in order to avoid premature convergence and reach approximate final solutions. This paper presents the gradual distributed real-coded genetic algorithms, a type of heterogeneous distributed real-coded genetic algorithms that apply a different crossover operator to each sub-population. Experimental results show that the proposals consistently outperform sequential real-coded genetic algorithms     

Asymptotic agreement and convergence of asynchronous stochastic algorithms

In this paper, we present results on the convergence and asymptotic agreement of a class of asynchronous stochastic distributed algorithms which are in general time-varying, memory-dependent, and not necessarily associated with the optimization of a common cost functional. We show that convergence and agreement can be reached by distributed learning and computation under a number of conditions, in which case a separation of fast and slow parts of the algorithm is possible, leading to a separation of the estimation part from the main algorithm.     

Performance analysis of direct N-body algorithms for astrophysical simulations on distributed systems

We discuss the performance of direct summation codes used in the simulation of astrophysical stellar systems on highly distributed architectures. These codes compute the gravitational interaction among stars in an exact way and have an O(N²) scaling with the number of particles. They can be applied to a variety of astrophysical problems, like the evolution of star clusters, the dynamics of black holes, the formation of planetary systems, and cosmological simulations. The simulation of realistic star clusters with sufficiently high accuracy cannot be performed on a single workstation but may be possible on parallel computers or grids. We have implemented two parallel schemes for a direct N-body code and we study their performance on general purpose parallel computers and large computational grids. We present the results of timing analyzes conducted on the different architectures and compare them with the predictions from theoretical models. We conclude that the simulation of star clusters with up to a million particles will be possible on large distributed computers in the next decade. Simulating entire galaxies however will in addition require new hybrid methods to speedup the calculation.     

Almost-sure convergence of adaptive algorithms by projections

A convergence proof is discussed for the normalized least-mean-square (LMS) algorithm for ergodic inputs. The approach is based on interpreting the algorithm as a sequence of relaxed projection operators by which the key contraction property is derived. The proof technique is strongly motivated by physical intuition, and provides additional insight into LMS-type algorithms under ergodic inputs. Embedded in the development is a slight generalization to a random time-varying gain parameter. This allows the incorporation of variations such as the LMS and signed LMS algorithms     

A proof of convergence for Ant algorithms

A proof of convergence for Ant algorithms is developed. Ant algorithms were modeled as branching random processes: the branching random walk and branching Wiener process to derive rates of birth and death of ant paths. Substitution is then carried out in birth–death processes which proves that a stable distribution is surely reached. This indicates that Ant algorithms converge with probability one. This analogy models Ant algorithms complexity parameters such as the number of cycles, the degrees of freedom of problem and the number of ants.