Box, Simon and Waterson, Ben
Machine learning in signalized junction control algorithms.
In, 43rd Annual Conference of the Universities' Transport Study Group (UTSG), Milton Keynes, GB,
05 - 07 Jan 2011.
Full text not available from this repository.
Machine learning techniques can be applied to develop signalized junction control algorithms that can learn control strategies from examples of good control and from experience. This paper discusses the conceptual differences between the conventional approach to signal control and the machine learning approach. An example is presented where a junction control agent was developed to learn strategies from a human expert. This learning junction agent uses localization probe data from vehicles and a system of bids to describe the state of the network. The junction agent learns from the human expert by employing a Neural Network to classify its bid space based on evidence of the human’s decision making.
Simulation experiments are used to evaluate the performance of learning junction agent and these show that the agent can outperform the High Bid signal control system both in terms of delay and in terms of equitability. The paper concludes with a discussion on how the approach described above can be extended to allow the junction control agent to learn from observational data and experience using reinforcement learning.
Actions (login required)