A machine learning approach for the dimensional reconstruction of the laminar artefacts of the MPPNB sites in the Southern Levant

The lithic industry of the Pre-Pottery B is based on the laminar reduction process that often ends with the production of tools made largely out of blades and secondly on bladelets. However, the artifacts state of conservation doesn’t always allow to have a complete dimensional set of data. The presence of broken laminar blanks due to the post-depositional processes is indeed high in many sites. In order to solve such problems, we propose a machine learning analysis as a neural network approach for an automatic reconstruction of missing dimensional data based on a standardised measurements‘ recording of laminar artefacts‘ dimensions. In Machine learning there are indeed two distinct models, one is called a supervised learning model and the other an unsupervised one, the latter is used to discover internal structures in datasets. In our case, we chose the supervised learning model, in which a feed-forward type neural network is trained with an input data vector that corresponds to a pre-established output vector, the algorithms used for learning are the Bayesian regularisation, the Levenberg-Marquardt and the Scaled conjugate gradient back-propagation. Particular attention has been paid to avoid underfitting and overfitting problems.