KFAE: Kalman Filter-Autoencoder Approach for Handling Faulty and Noisy Environments

Reinforcement Learning (RL) is widely applied in robotics, autonomous systems, and network optimization; however, it struggles with sensor noise and faulty actions, leading to instability. This paper introduces the Kalman Filter Autoencoder (KFAE), a novel approach that integrates Proximal Policy Optimization (PPO) with an Autoencoder and a Kalman Filter to denoise observations and correct action faults, thereby enhancing learning efficiency in noisy environments. KFAE stabilizes learning, improves decision reliability, and significantly outperforms the Noisy Environment (NE), achieving a 131.5% improvement in episode reward mean and an 184.1% increase in steps till collision. Additionally, KFAE closely approximates the Default Environment (DE) with minimal deviations, validating its effectiveness as a robust, fault-tolerant RL framework.