MUMBAI, India, Aug. 10 -- Intellectual Property India has published a patent application (202621072877 A) filed by Symbiosis International Deemed University on June 11, 2026, for Decentralized Multi-Agent Deep Reinforcement Learning System With Hybrid Reward For Autonomous Urban Traffic Signal Optimization.
Inventors include Pranav Lakhe; Parth Dhoke; Parth Vishnu; and Dr. Gagandeep Kaur.
The application for the patent was published on July 31, 2026, under issue no. 31/2026.
Abstract: ABSTRACT DECENTRALIZED MULTI-AGENT DEEP REINFORCEMENT LEARNING SYSTEM WITH HYBRID REWARD FOR AUTONOMOUS URBAN TRAFFIC SIGNAL OPTIMIZATION The present invention discloses a decentralized multi-agent reinforcement learning system (100) for autonomous urban traffic signal optimization. The system (100) comprises a traffic sensor array (110), a state observation module (120) with dynamic zero-padding, a plurality of independent Deep Q-Network agents (130) each assigned to a respective intersection, a hybrid reward computation module (140) combining a throughput maximization term with a squared-sum variance penalty term for equitable load distribution, a two-phase training and convergence module (150) with progressive learning rate decay and checkpoint recovery, an environment simulation interface (160), and an evaluation and metrics module (170). Each agent (130) independently observes local lane-level queue states and selects optimal signal phase transitions without inter-agent communication. The hybrid reward actively suppresses the queue-transfer pathology by superlinearly penalizing per-lane queue imbalances. Empirical evaluation on the Cologne8 benchmark demonstrates a 95.1% reduction in average waiting time, 81.1% reduction in maximum wait spikes. [
Disclaimer: Curated by HT Syndication.
