Multiple Objective Optimization for Fully Adaptive Active Sonar

Conventional active sonar systems rely on the operator to decide how to best allocate resources to detect and track targets. This may place impractical demands on the operator, particularly when multiple competing objectives are present. A fully adaptive active sonar learns the dynamics of an environment by looking for information on signals of interest and then adapts the system parameters to achieve objectives specified by the user. Traditionally in active sonar, detection and tracking of targets are two common objectives that are carried out non-simultaneously. This paper demonstrates an operator-free simplified intelligent active sonar system capable of simultaneous detection and tracking by allocating pings using a Reinforcement Learning (RL) algorithm. The system uses the popular state-action-reward-state-action RL agent combined with a modified reward metric inspired by multiple objective optimization. Initial evaluation of the approach shows that the intelligent active sonar can swiftly adapt to changes in the dynamic environment when tracking and detecting targets.