Trade-Offs in Navigation Problems Using Value-Based Methods

AI, Journal Year: 2025, Volume and Issue: 6(3), P. 53 - 53

Published: March 10, 2025

Deep Q-Networks (DQNs) have shown remarkable results over the last decade in scenarios ranging from simple 2D fully observable short episodes to partially observable, graphically intensive, and complex tasks. However, base architecture of a vanilla DQN presents several shortcomings, some which were mitigated by new variants focusing on increased stability, faster convergence, time dependencies. These additions, other hand, bring costs terms required memory lengthier training times. In this paper, we analyze performance state-of-the-art families mission created Minecraft try determine optimal for such problem classes cost accuracy. To best our knowledge, analyzed methods not been tested same scenario before, hence more in-depth comparison is understand real improvement they provide better. This manuscript also offers detailed overview methods, together with heuristics metrics registered during proposed mission, allowing researchers select better-suited models solving future problems. Our experiments show that Double networks are capable handling gracefully while maintaining low hardware footprint, Recurrent DQNs can be good candidate even when resources must restricted, double-dueling well-performing middle ground their performance.

Language: Английский

A New Hybrid Reinforcement Learning with Artificial Potential Field Method for UAV Target Search

Sensors, Journal Year: 2025, Volume and Issue: 25(9), P. 2796 - 2796

Published: April 29, 2025

Autonomous navigation and target search for unmanned aerial vehicles (UAVs) have extensive application potential in rescue, surveillance, environmental monitoring. Reinforcement learning (RL) has demonstrated excellent performance real-time UAV through dynamic optimization of decision-making strategies, but its large-scale environments obstacle avoidance is still limited by slow convergence low computational efficiency. To address this issue, a hybrid framework combining RL artificial field (APF) proposed to improve the algorithm. Firstly, task scenario training environment are constructed. Secondly, integrated with APF form that combines global local strategies. Thirdly, compared standalone algorithms analysis their differences. The experimental results demonstrate method significantly outperforms terms efficiency performance. Specifically, SAC-APF achieves 161% improvement success rate baseline SAC model, increasing from 0.282 0.736 scenarios.

Language: Английский