Online optimization of replacement policies using learning automata

• Published in: International journal of systems science 2008-03, Vol.39 (3), p.237-249
• Main Authors: Ikonen, E., Najim, K.
• Format: Article
• Language: English
• Subjects: Algorithms; Applied sciences; Discrete optimization; Exact sciences and technology; Industrial metrology. Testing; Inventory control, production control. Distribution; Learning; Mathematical optimization; Mechanical engineering. Machine design; Operational research and scientific management; Operational research. Management science; Preventive maintenance; Reliability theory. Replacement problems; Stochastic optimization
• ISSN: 0020-7721; 1464-5319
• DOI: 10.1080/00207720701750790

A global optimization algorithm operating online in a stochastic multi-teacher environment is suggested. An application example introduces a new perspective for solving some optimization problems dealing with reliability. First, a hybrid scheme combining reinforcement-based learning automata and confidence probabilistics is developed for a single-teacher environment. The scheme is able to find the optimal solution with high confidence, yet providing a sequence of search actions that converge to the minimal loss. In addition, the suggested approach provides an on-line measure of the confidence to the current solution. Second, a multi-teacher environment is considered. A simple application of a database enables any single-teacher reinforcement algorithm to be used for updating the learning automaton action probability distribution. Two alternative approaches are suggested, where the former provides superior performance in terms of confidence and loss; the latter is able to deal with dependencies between the cost and the duration of the evaluation of the cost function. The performance of the learning schemes is studied in simulations on maintenance optimization, where an accumulated number of failures is optimized online for a deteriorating production system with preventive maintenance. The simulations indicate superior performance of the hybrid scheme. A significant speed-up is observed by taking advantage of information from processes running online in parallel, thus making the learning automata approach a much more feasible approach for solving engineering problems of practical interest.