On The Estimation Of Initial Conditions Of ANeural Network For Resource Leveling

Abstract

A procedure for estimating the Lagrange multipliers, suitable for a neural network (NN) model for construction resource leveling (RL), is presented. The procedure uses a modi ed variable-reduction technique, in conjunction with some helpful suggestions on how to choose the initial values of the Lagrange multipliers. The model has been previously developed by mapping a formulation of the RL problem as an augmented Lagrangian multiplier (ALM) optimization, onto an arti cial neural network (ANN) architecture, employing a Hop eld-con guration of NN. In order to ensure the convergence of the NN model, a good estimate of the initial values of the Lagrange multipliers is needed. First, a non-singular decomposition of the constraint matrix is constructed, by taking into account at least one output variable for each noncritical activity. Then, the matrices corresponding to the cost function are similarly partitioned. Finally, rst-order multiplier estimates are used as an initial estimate of the Lagrange multipliers. An experimental veri cation of the proposed procedure is also provided.