Comparison of ANN Classifier to the Neuro-Fuzzy System for Collusion Detection in the Tender Procedures of Road Construction Sector

Abstract

As the contracts in the road construction sector in Poland are usually of extremely high value and financed from the state budget, the tender procedures should not allow for the non-concurrent behaviours of tender participants. Otherwise, the clients' losses will be of high value too. A database comprising hundreds of bidding procedures in the road construction industry in Poland has been developed. It includes the tender's participants, the locations of the roads sections, bids' values, the winners, and the types of roads. Every procedure has been evaluated and assigned to the set with a given level of collusion occurrence probability. The evaluation has required the analysis and the transformation of-described in the literature-collusive types of behaviours to the parameters of procedures that can be shown as numbers or ranks. Four criteria of a collusion threatened contracts have been chosen and applied for evaluation. Then, two methods of machine learning were applied. The first method was to train an artificial neural network (ANN) to classify the procedures to the aforementioned sets. The other method was to utilize artificial neural networks predictive capabilities enriched by the fuzzy sets theory. The multiple output from ANN was defined as membership function values. The use of the fuzzy sets theory-the process of defuzzification-helps to classify the tender procedures to the sets of different level of risk (of collusion appearance). The results achieved in these two separate processes are compared and discussed. The created tool can be applied for the future tender procedures as a pre-test of a collusion appearance.