A Custom Loss Function for Machine Learning-based Resource Allocation Policies

In this contribution, we present a new approach to minimize the outage probability of a machine learning-based (ML-based) resource allocation system by exercising a novel loss function. We demonstrate our approach on a single-user multi-resource greedy allocation system equipped with an ML model. Other resource allocation policies are apt for use though this is not the focus of our work. The ML model’s task is to assign resources to the user while minimizing outages. The critical issue is that this ML model has no access to any future channel state information but relies on its historical state to make future resource outage predictions. We first present outage probability expressions for this system. Following this, a custom loss function is presented that closely approximates the outage probability while also satisfying fundamental differentiable properties. Our experimental findings demonstrate that this loss function (i) is capable of effectively training our ML model in the presence of considerable biases in the training dataset, and (ii) significantly and invariably outperforms two commonly used loss functions, namely, the binary cross entropy (BCE) and mean squared error (MSE), in practical cases achieving improvements of two order of magnitude.