A neurodynamic model of a high-dimensional vector associative memory for
road traffic control optimization
- Dimitrios Lolas ,
- Cristian Axenie
Abstract
The manuscript introduces a novel perspective and approach to tackling
traffic control. Bypassing the need for computationally expensive
constrained optimization of spatiotemporal cues describing the road
traffic state, the system exploits the causal relation between traffic
light green light timing and the flow of cars. This way the system can
store traffic contexts as memories used to simply recall plausible green
time timings matching the flow of cars. This behavior amounts to an
autoassociative memory, efficiently implemented in spiking neural
networks, that provides a good trade-off between execution time and
accuracy. We believe the approach has very high potential in real-world
deployments. Our initial results on four real datasets demonstrate the
benefits that the approach has and, of course, the lightweight and
efficient computation steps and learning paradigm. Our goal is to raise
awareness in the traffic engineering community on how neural associative
memories can be a suitable candidate for traffic control, a problem
without straightforward solutions. The flexibility in representing
traffic data through vectors, causal learning of memories, and fast
recall provide outstanding benefits demonstrated through our
experiments. We hope the work will contribute to both the neural network
community, through a novel autoassociative memory system using
high-dimensional vectors, and the traffic engineering community, through
a novel solution to traffic control.