Computationally efficient neuro-dynamic programming approximation method for the capacitated re-entrant line scheduling problem

Abstract

This paper presents a computationally efficient neuro-dynamic programming approximation method for the capacitated re-entrant line scheduling problem by reducing the number of feature functions. The method is based on a statistical assessment of the significance of the various feature functions. This assessment can be made by combining the weighted principal components with a thresholding algorithm. The efficacy of the new feature functions selected is tested by numerical experiments. The results indicate that the feature selection method presented here can extract a small number of significant features with the potential capability of providing a compact representation of the target value function in a neuro-dynamic programming framework. Moreover, the linear parametric architecture considered holds considerable promise as a way to provide effective and computationally efficient approximations for an optimal scheduling policy that consistently outperforms the heuristics typically employed.

Keywords:

• capacitated re-entrant line
• data mining
• feature selection
• scheduling
• neuro-dynamic programming
• principal component analysis

Acknowledgement

This work was supported by Grant Nos. 2009-0070818 and 2010-0003811 from the National Research Foundation of Korea and Brain Korea 21 (Network Enterprise).

Notes

Note

1. For the case of configuration 1, the reported percent errors are non-zero as in Table 4, which results from the randomising nature of the applied policy.