Unsupervised machine-learning rule-based

RA1
University: Politecnico di Torino
Environment – Decarbonisation, Sustainability and Energy Efficiency
category: Road
N. project: L1-86
Members: Qinrui Tang
RA3
University: Technische Universität Braunschweig
Urban and Long-Distance People Mobility – Systems and Services
Unsupervised machine-learning
rule-based controller for dual-mode
parallel HEVs
Minimization of travel time in
signalized networks by prohibiting
left turns
In this project machine-learning has been
applied to generate a rule-based control
strategy for a parallel diesel hybrid electric
vehicle.
The control strategy selects the transmission
gear and manages the power-split between the
engine and the battery and aims at maximizing
the fuel economy with some constraints to
pollutant emissions and battery life depletion.
Instead of adopting heuristic techniques that
may lack of performance, rule-based controller
has been trained in this case with machinelearning. A clustering algorithm has been coded
to generate a 3D mesh for the vehicle velocity
and power, as well as the battery state of
charge. Genetic algorithms have been applied
to generate the optimal rule, which contains
the powertrain control action to take for each
cluster of the mesh. Specific features have been
The Direct left turn at intersections
increases delays and causes more accidents in
certain conditions. Prohibiting left turns could
increase the capacity of the road and reduce
the delays.
This research aims to analyze which left turns
should be prohibited to increase efficiency in
an urban network. Stochastic User Equilibrium
(SUE), with left turn prohibition, is studied in
this paper, by minimizing the travel time of
the network. The left turn prohibition solution,
the corresponding cycle time and green times
are determined. A left turn prohibition filtering
algorithm (LTPFA) is developed in this paper.
The core of the algorithm is that the small left
turn flows have priority to be prohibited. By
testing the LTPFA in an artificial network, the
LTPFA succeeds to find optimal or near-optimal
result of travel time in shorter time compared
extracted from a set of driving scenarios.
A dataset of driving missions as inputs and
of control rules as outputs has been created. The
vehicle control unit then receives the mission
features at the beginning of the trip and selects
the rule. During the trip, the cluster associated
to the instantaneous vehicle operating condition
is identified, and the control action is extracted
from the corresponding rule. The power-train
components are therefore actuated to drive the
vehicle.
with enumerating all left turn prohibition
combinations. As the total traffic demands
increase, the traffic network can benefit more
from the left turn prohibition.
However, when the traffic network is too
congested, the travel time reduction goes
down because no more green time can be
accommodated. An efficient left turn prohibition
decision would be applied in the demand around
the capacity. The phase sequence and offset
optimization are not included in this report.
Road
Members: Mattia Venditti
57
student
category: Road
N. project: L1-76