Hatem Ahriz

Solving Coarse-grained DisCSPs with Multi-DisPeL and DisBO-wd

We present multi-DisPeL, a penalty-based local search distributed algorithm which is able to solve coarse-grained distributed constraint satisfaction problems (DisCSPs) efficiently. Multi-DisPeL uses penalties on values in order to escape local optima during problem solving rather than the popular weights on constraints. We also introduce DisBO-wd, a stochastic algorithm based on DisBO (distributed breakout) which includes a weight decay mechanism. We compare multi-DisPeL and DisBO-wd with other algorithms and show, empirically, that they are more efficient and at least as effective as state of the art algorithms in some problem classes.

Building Models through Formal Specification

Over the past years, a number of increasingly expressive languages for modelling constraint and optimisation problems have evolved. In developing a strategy to ease the complexity of building models for constraint and optimisation problems, we have asked ourselves whether, for modelling purposes, it is really necessary to introduce more new languages and notations. We have analyzed several emerging languages and formal notations and found (to our surprise) that the already existing Z notation, although not previously used in this context, proves to a high degree expressive, adaptable, and useful for the construction of problem models. To substantiate these claims, we have both compiled a large number of constraint and optimisation problems as formal Z specifications and translated models from a variety of constraint languages into Z. The results are available as an online library of model specifications, which we make openly available to the modelling community.

Redesign Knowledge Analysis, Representation and Reuse

DEKLARE is a framework which enables the elicitation, representation and reuse of redesign knowledge. In this paper we present MAKUR, an extension to DEKLARE which allows the capture and management of more redesign knowledge, thus contributing towards a better and faster redesign of families of products.

A Hybrid Approach to Distributed Constraint Satisfaction

We present a hybrid approach to Distributed Constraint Satisfaction which combines incomplete, fast, penalty-based local search with complete, slower systematic search. Thus, we propose the hybrid algorithm PenDHyb where the distributed local search algorithm DisPeL is run for a very small amount of time in order to learn about the difficult areas of the problem from the penalty counts imposed during its problem-solving. This knowledge is then used to guide the systematic search algorithm SynCBJ. Extensive empirical results in several problem classes indicate that PenDHyb is effective for large problems.

Multi-Hyb: A Hybrid Algorithm for Solving DisCSPs with Complex Local Problems

A coarse-grained Distributed Constraint Satisfaction Problem (DisCSP) is a constraint problem where several agents, each responsible for solving one part (a complex local problem), cooperate to determine an overall solution. Thus, agents solve the overall problem by finding a solution to their complex local problem which is compatible with the solutions proposed by other agents for their own local problems. Several approaches to solving DisCSPs have been devised and can be classified as systematic search and local search techniques. We present Multi-Hyb, a two-phase hybrid algorithm for solving coarse-grained DisCSPs which uses both systematic and local search during problem solving. Phase 1 generates key partial solutions to the global problem using systematic search. Concurrently, a penalty-based local search algorithm attempts to find a global solution to the problem using these partial solutions. If a global solution is not found in phase 1, the information learnt from phase 1 is used to inform the search carried out during the next phase. Phase two runs a systematic search algorithm on complex variables guided by the following knowledge obtained in phase 1: (i) partial solutions and; (ii) complex local problems which appear more difficult to satisfy. Experimental evaluation demonstrates that Multi-Hyb is competitive in several problem classes in terms of: (i) the communication cost and (ii) the computational effort needed.

DynABT: Dynamic Asynchronous Backtracking for Dynamic DisCSPs

Constraint Satisfaction has been widely used to model static combinatorial problems. However, many AI problems are dynamic and take place in a distributed environment, i.e. the problems are distributed over a number of agents and change over time. Dynamic Distributed Constraint Satisfaction Problems (DDisCSP) [1] are an emerging field for the resolution of problems that are dynamic and distributed in nature. In this paper, we propose DynABT, a new Asynchronous algorithm for DDisCSPs which combines solution and reasoning reuse i.e. it handles problem changes by modifying the existing solution while re-using knowledge gained from solving the original (unchanged) problem. The benefits obtained from this approach are two-fold: (i) new solutions are obtained at a lesser cost and; (ii) resulting solutions are stable i.e. close to previous solutions. DynABT has been empirically evaluated on problems of varying difficulty and several degrees of changes has been found to be competitive for the problem classes tested.

CSP - there is more than one way to model it.

In this paper, we present an approach for conceptual modelling of constraint satisfaction problems (CSP). The main objective is to achieve a similarly high degree of modelling support for constraint problems as it is already available in other disciplines. The approach uses diagrams as operational basis for the development of CSP models. To facilitate a broader scope, the use of available mainstream modelling languages is adapted. In particular, the structural aspects of the problem are visually expressed in UML, complemented by a textual representation of relations and constraints in OCL. A case study illustrates the expositions and deployment of the approach.

Escaping local optima: constraint weights vs. value penalties.

Constraint Satisfaction Problems can be solved using either iterative improvement or constructive search approaches. Iterative improvement techniques converge quicker than the constructive search techniques on large problems, but they have a propensity to converge to local optima. Therefore, a key research topic on iterative improvement search is the development of effective techniques for escaping local optima, most of which are based on increasing the weights attached to violated constraints. An alternative approach is to attach penalties to the individual variable values participating in a constraint violation. We compare both approaches and show that the penalty-based technique has a more dramatic effect on the cost landscape, leading to a higher ability to escape local optima.

A Synergy of Modelling for Constraint Problems

Formulating and solving constraint problems requires both mathematical modelling as well as software development skills. Rather than reconciling these demands at implementation level, we have introduced a separate modelling layer in our research that abstracts away from low-level representation and implementation issues. We found that this benefits both the mathematical modelling and the software development aspects of constraint programming, leading to an efficient synergy of these activities. In this paper, we report on experiences and advances with our UML-based modelling approach.

Plan recommendation for well engineering

Good project planning provides the basis for successful offshore well drilling projects. In this domain, planning occurs in two phases: an onshore phase develops a project plan; and an offshore phase implements the plan and tracks progress. The Performance Tracker applies a case-based reasoning approach to support the reuse of project plans. Cases comprise problem parts that store project initiation data, and solution parts that record the tasks and subtasks of actual plans. An initial evaluation shows that nearest neighbour retrieval identifies projects in which the retrieved tasks and subtasks are relevant for the new project. The Performance Tracker can be viewed as a recommender system in which recommendations are plans. Thus the data that is routinely captured as part of the performance tracking during offshore implementation is utilised as experiences.