Collision prediction- based genetic network programming -reinforcement learning for mobile robot adaptive navigation in unknown dynamic environments

The autonomous navigation of a Mobile Robot (MR) in unknown environments populated by abundance of static and dynamic obstacles with a moving target have tremendous importance in real time applications. The ability of an MR to navigate safely, smoothly, and quickly in such environment is crucial. Current researches are focused on investigating these complex features in static or point-to-point dynamic environments. On the other hand, the salient downside of Q-Learning such as curse of dimensionality (CoD) is aggravated in complex environments. The objectives of this thesis is to address the issue of Adaptive Reinforcement Learning (RL) approaches in order to meet the requirements of MR navigation. Moreover, it aims to tackle CoD problem of Q-Learning (QL) to be suitable for complex applications. For this purpose, two genetic network programming with RL (GNP-RL) designs are proposed. The first design is based on obstacle target correlation (OTC) environment representation and called OTC-GNP-RL. This provides a perception of the current environment states. The second design is based on the proposed collision prediction (CP) environment representation and called CPGNP- RL. This representation is designed to provide collision prediction between MR and an obstacle, as well as the perception of current surrounded environment. Besides, it could represent an environment with compact state space and requires ones to measure positions only. Furthermore, the combination of CP and QL (CPQL) can overcome the downside of the CoD problem and improve navigation features.A simulation is used for evaluating the performance of the proposed approaches. The results show that the superiority of the proposed approaches in terms of the features of MR navigation, where all these features are taken under the design consideration of each proposed approach. Through the evaluation, CPQL, CP-GNP-RL, and OTCGNP- RL provide significant improvements in terms of safety (7.917%), smooth path (71.776%), and speed (10.89%), respectively, compared with two state-of-arts approaches, i.e. OTC based Q-learning and artificial potential field. In addition, the learning analysis of CPQL shows its efficiency and superiority in terms of learning convergence and safe navigation. Hence, the proposed approaches prove their authenticity and suitability for navigation in complex and dynamic environments.

Preview Text FK 2018 40 - IR.pdf Download (2MB) | Preview

Actions (login required)

View Item