Article
Deep Reinforcement Learning for Autonomous SONAR Port Monitoring (en)
* Presenting author
Abstract:
The use of MIMO-SONAR systems to autonomously monitor a port environment requires a robust control of the system parametrization and its adaption to changing environmen-tal conditions in real-time. Deep reinforcement learning can be used to implement a con-trol-assisting artificial intelligence (AI) which adapts the system parametrization in rela-tion to the observed environment scans. Through the design of a reward function, the controlling agent can learn the fulfillment of given sub-goals, such as the evaluation of security risks and the management of limited computational and energy resources. During training, the agent explores the unknown environmental dynamics in a trial-and-error fashion and improves its policy by exploiting the gathered knowledge of agent-environment interactions. By retrospective analysis of the chosen system parametrization and the resulting scan observations, the agent learns to adapt its monitoring strategy to fulfill the main goal of reliably detecting unwanted intruders inside of the port. This work presents the design of the reward function, the training architecture, and the perfor-mance evaluation of the trained deep reinforcement learning agent for a digital port envi-ronment.