Enhancement of Deep Deterministic Policy Gradient for Multi-Agent Robotics in Virtual Simulations

In this research, robots act as multi-agents for pathplanning techniques in unpredictable environments. By using the reinforcement learning method, the target point can be reached most efficiently from the starting point. This paper first introduces the Q-learning algorithm and deep Q-networks, which are implemented to improve the Multi-Agent Deep Deterministic Policy Gradient (IMADDPG). A new algorithm, namely IMADDPG, is proposed in this work based on the deep deterministic policy gradient, utilizing the centralized critic and actor decentralized (CCAD) framework. The benefits and drawbacks of guiding the robot with IMADDPG reinforcement learning techniques are discussed under both static and dynamic obstacle conditions. By employing a reinforcement learning technique, the robot can make intelligent decisions, adapt to changes in its surroundings, and learn from its experiences. To determine the potential of this method for practical applications, the effectiveness and performance of the proposed algorithm are analyzed. The ultimate goal of this work is to create a reliable and efficient robot navigation system capable of navigating effectively in unknown environments. The simulation results of the proposed algorithm are compared with MADDPG, DDPG, D3PG, and DQN, using training episodes up to 2000. The findings reveal a success rate of 96% and a failure rate of 1.2%.