A-DESIGN - AGENT BASED ADAPTIVE CONCEPTUAL DESIGN

INTRODUCTION | ITERATIVE SEARCH PROCESS | MULTIOBJECT DESIGN SELECTION | MULTI AGENT ARCHITECTURE | FUNCTIONAL REPRESENTATION | A DESIGN AS SEARCH STRATEGY | TEST RESULTS | DISCUSSION AND CONCLUDING REMARKS

DISCUSSION AND CONCLUDING REMARKS

This paper has introduced a new design generation theory known as A-Design that, through its implementation has shown that computer tools can adapt to change as well as play a part in the conceptual phase of the design process. A-Design is an agent-based and adaptive strategy for performing conceptual engineering design. The methodology has four distinct subsystems: an agent architecture, a multi-objective design selection scheme, a functional representation for electro-mechanical systems, and an iterative-based algorithm for evolving optimally directed design states. According to the theory, each of these subsystems plays an important role in enabling the production of creative and adaptive designs. The A-Design theory presented here provides a foundation for computer tools that can be developed to aid industries in meeting fast design cycles, in choosing ideal components from the overwhelming number of suppliers, and in meeting the variety of consumer demands.

Other search strategies or optimization techniques that solve problems through numerous permutations often require the existence of a well-defined set of variables prior to execution. However, defining these variables is often part of the conceptual design problem at hand. Knowledge based techniques can often overcome problems in ill-defined conceptual design, however these approaches often do not explore the vast space of conceptual designs as is done in stochastic search strategies. In fact, the conceptual design space is larger than the somewhat constrained parameter optimization space addressed by stochastic strategies, thereby suggesting that a mechanism for generating conceptual design explores as many alternatives as possible in order to find successful design alternatives. It is for this reason that the A-Design approach presented here combines aspects of both stochastic optimization and knowledge based design strategies. The knowledge-based strategy is contained within software agents which interact within an iterative algorithm to search the design space in a stochastically guided manner for the solutions that best meet user specifications.

Initial test examples help to demonstrate the effectiveness of the algorithm’s adaptability and search success. The Manhattan Transfer test example finds solutions for a specific user preference and can quickly adjust when a change in preference occurs. The numerical optimization test example pits the A-Design theory against a traditional SQP algorithm, and although the SQP method runs faster than A-Design, it often gets stuck in local optima while A-design is able to more consistently find the global optimum. Agent interaction and pareto design selection in A-Design allow solutions to be searched in the non-monotonic and multi-modal spaces of the test examples. For a conceptual design problem, the A-Design system has also been experimentally shown to produce successful designs, as demonstrated by the weighing machine example. Furthermore, experiments with this problem have validated the theory’s separation of designs into pareto, good, and poor subsets as an effective way to both optimize objectives and retain flexibility. The “recessive? characteristics preserved in the array of alternative designs provide adaptiveness to changing conditions.

The electro-mechanical design system is capable of not only producing real designs, but of producing a diversity of solutions that are representable and configurable by the functional representation scheme and agent architecture. The agents of the A-Design system contain very little information specific to electro-mechanical design. Agents are based on simple reasoning strategies such as tree search and pattern matching. In order to keep agents as general as possible for a wide variety of design problems, the knowledge of how components influence one another in a design has been placed in the embodiment representations. This generality allows for a variety of embodiments such as multi-input/multi-output components, electrical components, as well as characteristically different mechanical components from the rotational, hydraulic, and translational domains.

The power of A-Design’s generality is due to the iterative approach to design generation. Since functionality of a given device is specified in part by the objectives that define the design problem, generated solutions can be compared on a common metric of how well they satisfy the design specifications. Initially, agents design relatively poor solutions, but as the design selection scheme isolates better alternatives and feedback is provided to the agents, the process improves designs to best meet the user’s defined functionality. For example in the weighing machine problem, accuracy of dial is an objective that partially defines the functionality of the device. Initially, agents may create solutions that do not even cause the output dial to rotate. As the process unfolds, solutions appear that better meet the accuracy objective, and feasible weighing machines are created. Future work with the electro-mechanical A-Design approach will accommodate dynamic analyses to allow the user to better describe a design problem through desired system dynamics such as the settling time of the dial in the weighing machine example.

The premise of this work is that design takes place within a dynamic environment. Thus far we have focused on changing user preferences for the various objectives describing a design problem. In current work, A-Design’s adaptability is being tested in cases where objectives not only change in importance but also are added or removed from the design problem. This enables the addition and deletion of design constraints when modeled as penalty functions and treated as objectives in the optimization process. Finally, our model also can readily incorporate changing component embodiments; as new technologies emerge or older ones are retired, components can be readily added and removed from the catalog. Future work will explore these and other means of modeling the changing environment in which A-Design is able to create useful and innovative design concepts.