Reinforcement Learning-Based Optimization in Autonomous Intelligent Systems
DOI:
https://doi.org/10.5281/zenodo.19152624Keywords:
reinforcement learning, utonomous systems, optimisation, model-based RL, world model, evolutionary strategy, sample efficiencyAbstract
Autonomous intelligent systems operating in complex, high-dimensional environments require optimisation strategies capable of learning efficient
policies without exhaustive prior knowledge of the environment dynamics. Reinforcement learning (RL) has demonstrated strong potential in this
domain, yet standard model-free algorithms suffer from low sample efficiency, reward sparsity, and instability in continuous action spaces. This study
introduces OptiRL, a model-based reinforcement learning framework that integrates a differentiable world model with a population-based
evolutionary strategy and a prioritised experience replay mechanism to achieve robust, sample-efficient optimisation in autonomous systems. OptiRL
was benchmarked across five autonomous control tasks: robotic arm manipulation (MuJoCo Reacher-v4), autonomous vehicle lane-change
optimisation (CARLA simulator), drone trajectory planning (AirSim), industrial process control (custom refinery environment, n = 240,000 episodes),
and multi-robot warehouse coordination (n = 48 agents). OptiRL achieved mean cumulative reward improvements of 38.4% over SAC, 44.7% over
PPO, and 29.3% over DreamerV3 across all five tasks. Sample efficiency improved by 52.1% on average, reaching equivalent performance in 47.9%
fewer environment interactions. These results demonstrate that world-model integration with evolutionary search provides a principled and effective
framework for autonomous system optimisation across diverse control domains.
