Load balancing is one of the central problems that have to be solved in parallel computation. Here, the problem of distributed, dynamic load balancing for massive parallelism is addressed. A new local method, which realizes a physical analogy to equilibrating liquids in multi-dimensional tori or hypercubes, is presented. It is especially suited for communication mechanisms with low set-up to transfer ratio occurring in tightly-coupled or SIMD systems. By successive shifting single load elements to the direct neighbors, the load is automatically transferred to lightly loaded processors. Compared to former methods, the proposed Liquid model has two main advantages. First, the task of load sharing is combined with the task of load balancing, where the former has priority. This property is valuable in many applications and important for highly dynamic load distribution. Second, the Liquid model has high efficiency. Asymptotically, it needs O(D . K . Ldiff ) load transfers to reach the balanced state in a D-dimensional torus with K processors per dimension and a maximum initial load difference of Ldiff . The Liquid model clearly outperforms an earlier load balancing approach, the nearest-neighbor-averaging. Besides a survey of related research, analytical results within a formal framework are derived. These results are validated by worst-case simulations in one-and two-dimensional tori with up to two thousand processors.
The paper presents a novel approach to parallel motion planning for robot manipulators in 3D workspaces. The approach is based on a randomized parallel search algorithm and focuses on solving the path planning problem for industrial robot arms working in a reasonably cluttered workspace. The path planning system works in the discretized configuration space which needs not to be represented explicitly. The parallel search is conducted by a number of rule-based sequential search processes, which work to nd a path connecting the initial configuration to the goal via a number of randomly generated subgoal configurations. Since the planning performs only on-line collision tests with proper proximity information without using pre-computed information, the approach is suitable for planning problems with multirobot or dynamic environments. The implementation has been carried out on the parallel virtual machine (PVM) of a cluster of SUN4 workstations and SGI machines. The experimental results have shown that the approach works well for a 6-dof robot arm in a reasonably cluttered environment, and that parallel computation increases the efficiency of motion planning significantly.
We present an approach to systematically describing case-based reasoning systems bydifferent kinds of criteria. One main requirement was the practical relevance of these criteria and their usability for real-life applications. We report on the results we achieved from a case study carried out in the INRECA1 Esprit project.
We present a novel approach to classification, based on a tight coupling of instancebased learning and a genetic algorithm. In contrast to the usual instance-based learning setting, we do not rely on (parts of) the given training set as the basis of a nearestneighbor classifier, but we try to employ artificially generated instances as concept prototypes. The extremely hard problem of finding an appropriate set of concept prototypes is tackled by a genetic search procedure with the classification accuracy on the given training set as evaluation criterion for the genetic fitness measure. Experiments with artificial datasets show that - due to the ability to find concise and accurate concept descriptions that contain few, but typical instances - this classification approach is considerably robust against noise, untypical training instances and irrelevant attributes. These favorable (theoretical) properties are corroborated using a number of hard real-world classification problems.
We present a similarity criterion based on feature weighting. Feature weights are recomputed dynamically according to the performance of cases during problem solving episodes. We will also present a novel algorithm to analyze and explain the performance of the retrieved cases and to determine the features whose weights need to be recomputed. We will perform experiments and show that the integration in a feature weighting model of our similarity criterion with our analysis algorithm improves the adaptability of the retrieved cases by converging to best weights for the features over a period of multiple problem solving episodes.
Planning for manufacturing workpieces is a complex task that requires the interaction of a domain-specific reasoner and a generic planning mechanism. In this paper we present an architecture for organizing the case base that is based on the information provided by a generic problem solver. A retrieval procedure is then presented that uses the information provided by the domain-specific reasoner in order to improve the accuracy of the cases retrieved. However, it is not realistic to suppose that the case retrieved will entirely fit into the new problem. We present a replay procedure to obtain a partial solution that replays not only the valid decisions taken for solving the case, but also justifications of rejected decisions made during the problem solving process. As a result, those completion alternatives of the partial solution are discarded that are already known to be invalid from the case.
We present an algorithm for completely replaying previous problem solving experiences for plan-space planners. In our approach not only the solution trace is replayed, but also the explanations of failed attempts made by the first-principle planner. In this way, the capability of refitting previous solutions into new problems is improved.
This paper addresses the role of abstraction in case-based reasoning. We develop a general framework for reusing cases at several levels of abstraction, which is particularly suited for describing and analyzing existing and designing new approaches of this kind. We show that in synthetic tasks (e.g. configuration, design, and planning), abstraction can be successfully used to improve the efficiency of similarity assessment, retrieval, and adaptation. Furthermore, a case-based planning system, called Paris, is described and analyzed in detail using this framework. An empirical study done with Paris demonstrates significant advantages concerning retrieval and adaptation efficiency as well as flexibility of adaptation. Finally, we show how other approaches from the literature can be classified according to the developed framework.
This paper is to present a new algorithm, called KNNcost, for learning feature weights for CBR systems used for classification. Unlike algorithms known so far, KNNcost considers the profits of a correct and the cost of a wrong decision. The need for this algorithm is motivated from two real-world applications, where cost and profits of decisions play a major role. We introduce a representation of accuracy, cost and profits of decisions and define the decision cost of a classification system. To compare accuracy optimization with cost optimization, we tested KNNacc against KNNcost. The first one optimizes classification accuracy with a conjugate gradient algorithm. The second one optimizes the decision cost of the CBR system, respecting cost and profits of the classifications. We present experiments with these two algorithms in a real application to demonstrate the usefulness of our approach.