Disertasi

Hybrid genetic algorithm and reinforcement learning for mobile robot path planning optimisation in dynamic environment / Dyah Lestari

Abstrak
Mobile robots are increasingly applied in industrial settings for efficient and safe task execution with path planning being one of the major challenges. Classic algorithms can handle path planning in static environments but often fail in dynamic spaces such as warehouses where obstacles can be placed randomly or moved frequently. Genetic Algorithms (GAs) have been explored for their global search capabilities but standard GAs face problems of slow convergence or suboptimal results due to fixed parameter settings and limited adaptation to environmental changes. This dissertation addresses these challenges through three stages of research. First a new GA model was developed with grid-based genetic representation to simulate robot path planning in environments such as warehouses with static obstacles. Simulations were conducted on six variations of start-point ndash destination-point combinations with random obstacles. Second a performance improvement method called Dynamic Crossover and Mutation Rate (DCMR) was introduced in GA (GA DCMR) where crossover and mutation rates are linearly adjusted across generations (0%-100% crossover 100%-0% mutation). This method can improve adaptability and eliminate the need for manual parameter adjustment. This method was tested in six scenarios as above. Third a hybrid GA and Reinforcement Learning (GA RL) algorithm was proposed to handle dynamic changes such as the addition removal and movement of obstacles. Obstacle changes are detected using the Bayesian Occupancy Grid (BOG) where GA perform global planning and Q-learning adjusts the path locally based on environmental feedback. Algorithm performance is evaluated using completeness path length optimality convergence rate and computational time. GA with a new genetic representation successfully found feasible paths in 95% of cases although the convergence rate slowed down as the initial-destination distance increased. GA DCMR achieved 100% completeness and increased the convergence speed by 89.48% compared to fixed parameter settings demonstrating its ability to balance exploration and exploitation. The hybrid GA RL algorithm further overcomes the limitations of GA in dynamic environments consistently detecting obstacle changes achieving 100% completeness generating near-optimal paths (94%-100%) achieving a convergence rate of 1 and converging in 3 seconds in real-time trials. Moreover RL helped guide GA to avoid local optima and enhanced convergence stability. Despite strong results in completeness optimality and convergence rate this study has limitations the obtained paths are not always the shortest experiments were limited to small grid-based simulations BOG may degrade under high-frequency dynamics the algorithm has not been tested on real robots and resource usage was not analysed. Future work will refine hybrid GA RL with local optimization test scalability in larger environments validate performance on physical robots explore advanced adaptive parameter control and evaluate memory and energy efficiency for embedded and mobile systems.

Informasi Detail

DDC

Rd 629.892 LES h

Prodi

Universitas Negeri Malang. Program Studi Teknik Elektro dan Informatika, 2025.

Deskripsi Fisik

xviii, 110 lembar : ilus. ; 30 cm.

Bahasa

No Reg

00278/RD/25

Edisi

Disertasi (Pascasarjana)--Universitas Negeri Malang. 2025

Subjek

1. ROBOTIKA - OPTIMASI PERENCANAAN JALUR
2. ALGORITMA KOMPUTER
3. ROBOTICS - PATH PLANNING OPTIMIZATION

Pembimbing

1. Dr.eng Siti Sendari, S.t., M.t.; 2. Ilham Ari Elbaith Zaeni, S.t., M.t., Ph.d

Lampiran Berkas