`

Case Study 4 : DIPSO - Distributed Problem Solving Environment

Overview of DIPSO.

DIPSO is a problem solving environment (PSE) which integrates modelling and optimisation services located at geographically distributed centres. Integration is supported through Grid technologies [1] and software components relating to system modelling and design search, exploration and analysis. These processes return high-quality design solutions and relevant design information to support conceptual design decision-making. 

An appropriate conceptual design problem has been provided by Systems Engineering and Assessment (SEA ) Ltd, Frome, UK. This comprises a parametric model of a remote-operated undersea vehicle. Design variables within this conceptual whole-system model include overall vehicle dimensions, characteristics of single or hybrid power sources, fuel and equipment loadings, etc The performance of a trial vehicle can be assessed across circa fifty possible missions. 

The initial distributed system architecture, modelled via a Triana workflow engine [2], is shown below and the functionality of each components is described. 

The Modeller currently comprises a number of ancillary data processing techniques such as clustering and principal component analysis plus a Radial Basis Function neural network [3] and standard statistical modelling software for the generation of response curves from incoming data sets passed into the system by the client. 

The Interrogator extracts information relating to design space characteristics from a parametric model. The basic prototype comprises Halton injection sequences [4] to identify well-distributed points within the design space. Standard hill-climbers then search from these points. Although initially simplistic, this investigation provides an indication of the number and distribution of local optima. For instance, if all hill-climbers converge to the same point the surface is likely to be uni-modal / monotonic and no further search is required. However, if all hill-climbers converge to differing points in a particular region of the space then local, non-linear search technique such as Tabu or Simulated Annealing is selected from the Optimiser to further search these regions. Alternatively if all hill-climbers converge to points well-distributed across the space a more global, non-linear search technique such as a genetic algorithm is introduced. Other investigatory algorithms are currently being included in the Interrogator to provide further information and support the introduction of alternative rules. These may, for instance, recommend the utilization of multiple search techniques to provide global and local search to overcome multiple local optima and / or discontinuity.

The Optimiser currently comprises three search and optimization algorithms: a genetic algorithm, a simulated annealing algorithm and a tabu search algorithm [5]. The intention here is to provide a library of local and global search and optimization procedures that can be combined to achieve robust, high-performing search, exploration and optimization systems. Other stochastic and more deterministic optimization algorithms will be integrated as needs arise. 

The Knowledge Repository provides a storage capability for background information submitted by the clients and emerging information appertaining to problem characteristics and solutions from the Interrogator and Optimiser. 

Application

The two main modules, the Interrogator and the Optimiser were initially developed for a distributed grid-based environment. These are now available as either individual stand-alone tools or coupled to provide a modelling and search/optimisation capability.
The Interrogator is suitable if there is an existing parametric model describing the search space, and can be used in a number of ways: 1) To determine whether the parametric model is valid, 2) to discover high-performance regions, and 3) to explore the characteristics of the surface. The Optimiser provides a library of search algorithms which can be used to explore a suitable search space for optimal solutions. The various algorithms have been designed so that they do not require complex parameter settings.

The two modules have been tested and proofed within a drug discovery environment with Evotec, and the performance has been assessed using a range of test functions and the SEA parametric model. These trials have demonstrated the capability of the system to return both high quality solutions and relevant design information about likely design space characteristics in real industrial scenarios.
References 

1. O. F. Rana, L. Pouchard, 2003, Agent Based Semantic Grids: Research Issues and Challenges. Journal of Parallel and Distributed Computing Practices.

2. S. Majithia, M. Shields, I. Taylor, I. Wang. 2004, Triana: A Graphical Web Service Composition and Execution Toolkit. IEEE International Conference on Web Services, San Diego.

3. S. Haykin, 1999, Neural Networks: A Comprehensive Foundation, Second Edition, Macmillan, New York.

4. Sobol, I. M. 1967. On the Distribution of Points in a Cube and the Approximate Evaluation of Integrals”. Computational Mathematics and Mathematical Physics, 7, (4), pp 784-802. 

5. I. C. Parmee, 2001, Evolutionary and Adaptive Computing in Engineering Design. Springer Verlag, London.