The Pittsburgh Youth Study found that 57 percent of juvenile delinquents continue to offend up to age 25. Juveniles who start offending before age 14 are likely to continue offending into early adulthood. As Erica Goode pointed out in her studies, this may be a reflection of the aggressive criminal justice and zero-tolerance policies for juvenile delinquents.

     In our study, we wanted to offer an alternative to the classical aggressive criminal sentences. Using a combination of public and private data, a Machine Learning algorithm suggests a personalised sentence for a juvenile criminal. The prediction takes into account the needs of each juvenile and tries to find an alternative solution to prison.

​

       The software was developed in Python, using the open sources libraries TensorFlow, PyTorch, Keras and Pandas. The data was obtained from a small database provided by the Faculty of Law of Maastricht University. The model was defined by six layers. It was Trained with stochastic gradient descent for optimization, provided by the SGD class, and  with categorical cross-entropy loss for multi-class classification as its loss function, provided by the CrossEntropyLoss class, both classes from PyTorch.

​

        The testing process was bounded by the limited data availabe. Therefore, more data would be necessary to conduct a proper testing process.