Browsing by Author "Hossain Ahamed"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Open Access
    RER: Recycled Experience Replay with Dual Memory Architecture for Path Planning of a Moving Target using Deep Reinforcement Learning
    (North South University, 2021) Samiya Kabir Youme; Hossain Ahamed; Towsif Alam Chowdhury; Sayeed Abid; Rusafa Binte Sohrawardi; Shahnewaz Siddique; 1711848042; 1711678042; 1712399042; 1711870042; 1711462042
    All over the world, SAR operations are carried out to assist people in life-threatening situations. To search for a person in a life-threatening situation, the use of Unmanned Aerial Vehicles (UAVs) has increased drastically for different search and rescue missions to find the person at the earliest over the past few years. As UAVs are getting cheaper with advanced features like high-resolution cameras and long-lasting batteries, these devices are being used for autonomous search and rescue operations in different types of terrains and environments. These autonomous devices use artificial intelligence methods such as deep reinforcement learning algorithms for finding the optimal path and tracking the target. For marine-based environments, the target is continuously drifting with the ocean current, which makes it quite difficult for the UAV to search for the lost victim. In this project, we have made a simulation of a custom 2D marine environment and developed a dual memory architecture for finding the optimal path of a moving target to improve the learning of a UAV. We have incorporated our algorithm into popular deep reinforcement learning algorithms and improved the performance of classical algorithms by using our recycled experience replay. The results delineate that with a simple dual memory structure, immense progress in stable learning behavior can be obtained. The main goal of this project is to enhance the performance of prevalent deep reinforcement learning algorithms and test their performance in a simulated marine environment.