Improving the Performance of AI Algorithms.
Abstract
The feasibility of improving the efficiency of AI(Artificial Intelligence) software using available systems and methodologies is addressed. By modeling program behavior as a series of concurrent problem solution systems, it is possible to isolate the inefficiencies inherent in the implementation scheme for those due to conceptual difficulties or inadequacies in the underlying physical system. The processing environment selected for the implementation of AI software effectively establishes a computational paradigm which shapes the development and ultimate performance of any program executing within it. Sequential environments view the underlying architecture as von Neuman and approach a problem in terms of the Turing Model of Computation, while applicative environments exemplify the recursion theory approach. Established optimization techniques are intimately tied to the computational model and cannot be transported from one environment to the other with ease or efficiency. Since some AI tasks are inherently sequential and others inherently recursive, non single processing system can facilitate uniformly optimum performance. The concept of 'environment spanning' is suggested as a means of maximizing program optimizability by allowing the assignment of subproblems individually to whatever processing system offers the best chance for automatic improvement. Three mechanisms for implementing spanned environments are presented: parallel environments, multitasked environments, and intersequenced sub-environment modules.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1987
- Accession Number
- ADA192848
Entities
People
- Cherri M. Pancake
Organizations
- Auburn University