Comparison of Effectiveness of Machine Learning Algorithms for Vehicle Path Prediction

In modern warfare, intercepting moving enemy targets such as tanks, aircraft, missiles, and drones plays a crucial role. These targets are either controlled by enemy personnel or by sophisticated electronic systems. Therefore, their movements are best characterized by random motion subject to certain physical laws. Predicting these motions is extremely complex and often requires continuous tracking through sophisticated radar equipment. Machine Learning algorithms, such as Artificial Neural Networks, have proven to be effective in learning many real world motions of vehicles on the roads and have been extensively used in the autonomous vehicles. Artificial Neural Networks use activation functions to determine the output of a model from the given observations. After training the model with appropriate activation function, the model can be used for predictions. In this process, the activation functions play a crucial role. Selecting the correct activation function is critical to the success of the model. This project simulates the moving enemy target using a BristleBot (a brush-head fitted with vibrating motor which generates vibrations in the bristles thus propelling the BristleBot) which moves on a flat surface. The motion of the BristleBot is digitized by recording the X-Y coordinates on the path it has taken from the beginning of the run to the end of the run. These runs are repeated and data from multiple runs is stored in a database. Using R Programming language, a neural network training algorithm is simulated where the activation function can be changed (slope-intercept linear function y = mx + b with various slopes and intercepts, quadratic function y = a x
²
+ bx + c with various a, b, and c values). The resulting models corresponding to each training session are compared with each other to find their similarity to the paths taken by the BristleBot. The effectiveness of these activation functions is then measured by the similarity score. The trained model (or the activation function) with best similarity score is then selected for predicting the future path of the BristleBot. This model then can be stored on a chip and interceptor vehicles can use it to predict the path and intercept the target. This project is a simulation to demonstrate the usefulness of the Machine Learning algorithms (especially, Neural Networks) to train the models and store them on a chip that can guide the autonomous drones and missiles where sophisticated radar and satellite equipment are not feasible to guide them more accurately. Small inexpensive drones can be equipped with these chips to predict the paths of moving targets. Swarming with such drones is more economical in intercepting the targets. The simulation results with BristleBot are analyzed and similarity scores are obtained for different functions. These results indicate a reasonable effectiveness of quadratic functions for path prediction. The poster describes the simulation, linear and quadratic functions and their similarity scores, and the further research.