Optimization of multi-agent traffic network system with Q-Learning-Tune fitness function
This study aims to explore the potential of implementing multi-agent-based Genetic Algorithm (GA) with interactive metamodel to acquire regular optimisation on dynamic characteristic of traffic flow. The idea is proposed in effort of accessing the functionality of the proposed algorithm to improve t...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/37658/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/37658/2/FULLTEXT.pdf https://eprints.ums.edu.my/id/eprint/37658/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This study aims to explore the potential of implementing multi-agent-based Genetic Algorithm (GA) with interactive metamodel to acquire regular optimisation on dynamic characteristic of traffic flow. The idea is proposed in effort of accessing the functionality of the proposed algorithm to improve the smoothness of traffic flow in a network. As such, level-of-service of traffic network could be improved by optimising the utilisation of network capacity while minimising the travel delay and vehicles in queue. Traditionally, the common practice is to identify a fixed timing plan profile via offline and assumed it as a "nominal optimised" for the actual traffic flow. Whether the traffic signal is fully optimised under various traffic conditions, fluctuations in traffic demand and numerous uncertainties due to driver's driving behaviour remaining as a challenging topic. Scholars have proposed artificial intelligence (AI) to be integrated into the signal control system to improve the adaptiveness of the control system. However, the evaluation function used in the AI is developed based on historical traffic data. This offiine predetermined evaluation function has limited the AI in exploring the stochastic and non-uniform traffic flow environment to search the optimum solution. Therefore, a notable fitness function with interactive metamodel for GA or known as improved GA is proposed. The dynamic environment causing the need of dynamic modelling for better dynamic optimisation will be catered via a specifically formulated interactive fitness function. The interactive metamodel is extracted using Q-Learning (QL) via online observing and learning of the outflow-inflow traffic characteristics. The improved GA is then embedded into the signal controller of every intersection or known as agent. Each agent has the autonomy in controlling their local intersection which are coordinated by a superior agent that has superiority in overwriting the local control decision if conflict occurs. The improved GA is tested using simulated grid traffic network model under various traffic scenarios. Results indicate the improved GA has improved 7.0 - 9.0 % in minimising the average delay as compared to the classical GA (without interactive metamodel). |
|---|