2015 Results

Further extensions of the energy-aware arterial bandwidth optimization problem has been devised. A thorough comparison of the proposed strategy with other algorithms existing in the literature has been performed. In particular, the presented solution has been proved to outperform not only the offset-based optimization proposed by Gomes [IEEE TITS 2015], but also known arterial bandwidth strategies, such as MAXBAND [TRB 1981].

Furthermore, the original stringent assumption of under-saturated traffic conditions, that is the assumption of a traffic demand that guarantees the existence of a green wave, has been relaxed. In particular, a study at different levels of traffic demand and congestion has been conducted to assess the performance of the proposed strategy in challenging scenarios. It has been demonstrated that it is possible to tune the weights of the optimization function in order to reach Pareto-optimal operation conditions, also at over-saturated traffic conditions.

A new comparison with the existing arterial bandwidth algorithms at this point has demonstrated that our extension to a demand-adaptive framework allows to further increase the gains and the benefits in terms of energy consumption, average travel time, number of stops and time wasted idling at traffic lights.

During the visit of Alain Kibangou at UC Berkeley, discussions with Gabriel Gomes and Cheng-Ju Wu (PhD student) took place according to traffic prediction issues. For traffic flow prediction, the UC Berkeley team has developed a boundary flow prediction method that combines the most recent traffic data with historical traffic data. For this purpose an autoregressive moving average with an exogenous input (ARMAX) model is estimated online with the most recent vehicle detector station data. Then a multiple-step-ahead predictor of traffic demand is obtained from the estimated ARMAX model by solving a corresponding Bezout equation for each predictor. At the INRIA side, instead of ARMAX, flow or travel time are model as random walk models whose statistical properties are learned from historical data. Based on clustering of these historical data, different models can be learnt. Then a Kalman filter can be designed for each cluster before fusing the different predictors.  It is rather difficult to state about the optimality of one approach with respect to the other. Both are based on learning a model, deterministic (ARMAX) or stochastic (random walk), from historical data. It is worth noting that while the traffic prediction topic is active at the INRIA side through the PhD thesis of Andres Ladino, it is currently on standby at the Berkeley side since the departure of Dr Thomas Schreiter from the team. However, lessons learnt from the two approaches can give insights to develop efficient method in particular for large scale traffic network. Indeed, for such networks, simple models that can be efficiently learnt online must be developed. ARMAX can be too complex while first order random walk model can be too simple to capture the whole dynamics of the traffic flow. New generation of model is then to be investigated with new people involved in this topic at UC Berkeley side.