ROBUST MULTI-AGENT PATH PLANNING ALGORITHM FOR TARGET LOCALIZATION IN AN UNKNOWN ENVIRONMENT
This study proposes a new approach to improve the robustness of the path planning algorithm based on received signal strength (RSS) by defining several variables as the state of Q-learning. Usually, there are two approaches to defining state Q-learning in RSS-based target localization problems, both...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/73149 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | This study proposes a new approach to improve the robustness of the path planning algorithm based on received signal strength (RSS) by defining several variables as the state of Q-learning. Usually, there are two approaches to defining state Q-learning in RSS-based target localization problems, both of which only use one variable as state Q-learning. Either approach is potentially difficult to achieve convergence. While other approaches are based on unrealistic assumptions and conflict with radio-frequency (RF) wave propagation characteristics. In this study, it will be proven that if these assumptions are not met, the performance of the methods proposed by other studies will not be able to converge. In addition, this research proposes a new approach using two variables as state Q-learning. This approach is proven to be able to outperform approaches that only use one variable as the Q-learning state in terms of speed to achieve convergence and robustness.
In addition, this study also proposes using the Boltzmann distribution to replace ?-greedy, which is commonly used as an exploratory-exploitation method in RSS-based target localization research. This aims to accelerate the learning process of the algorithm so that the algorithm does not require many training episodes to achieve convergence. The simulation results show that using the Boltzmann distribution can reduce the steps needed for an agent to reach the target by up to 70% in the first ten training episodes.
Lastly, this research uses a cooperative multi-agent approach to solve the target localization problem based on RSS. In the approach taken, it is proposed to combine the Q-tables owned by each agent while still prioritizing the agent's own Q-value. The simulation results show that using this algorithm can reduce the number of steps by up to 18% compared to independent learners in the first ten training episodes
|
|---|