Applying Novel Adaptive Activation Function Theory for Launch Acceptability Region Estimation with Neural Networks in Constrained Hardware Environments: Performance Comparison

In this paper, we propose to apply Adaptive Activation Functions (AAF) theory and custom loss functions to classical Multi-layer Feedforward Neural Networks to improve their performance and reduce their model parameter size for solving the Launch Acceptability Region (LAR) problem in a specific guidance mode for a missile. We investigate and compare neural networks’ performance, including AAF, with other common methods such as Multivariate Adaptive Regression Spline (MARS) and lookup table interpolation for solving complex LAR prediction problems for deployment environments with limited hardware resources. We also introduce a novel loss function for training neural networks specific to the solution of the LAR problem. We demonstrate that using AAF in feed-forward neural networks as universal function approximators gives more accurate estimations with fewer model parameters. Experiment results show that our proposed method, using AAF in neural network layers, achieved lower Mean Absolute Error (MAE) and Mean Squared Error (MSE) values, as well as higher Intersection over Union (IoU) scores against classical networks and other regression and lookup-table-based methods for a specific guidance mode. Our experiments also elaborate that adding extra trainable parameters in the activation function equations for each layer helped to extend its learning capability and generalization of the networks, and significantly reduced model parameter size for deployment amongst other algorithms. As a result, we observe that this solution is better suited for resource-constrained hardware environments for deployment of the developed models, and AAF theory successfully applies to regression-based LAR prediction task in addition to physics aware and physics informed neural networks.