Simulation-based traffic management for autonomous and connected vehicles

Simulation-based traffic

management for autonomous and

connected vehicles

Paweł Gora

ITS Kraków, 3-4.12.2015

Faculty of Mathematics, Informatics and Mechanics

University of Warsaw

Axioms

● _{Vehicles may communicate with each other (V2V) and with infrastructure (V2I, I2V)} ● _{Car’s onboard computer may „know”:}

− _{static components of the road network} − _{precise location and speed of the car}

● _{GNSS data (GPS, Galileo + EGNOS)} ● _{sensors: lidar, laser, camera, …}

− _{intended destination and route of the car}

− _{similar dynamic attributes of other vehicles (thanks to V2V, V2I, I2V) and bikes /} pedestrians from the vicinity (sensors, mobile devices)

● _{Traffic management center may know the static road network structure, locations, speeds,} intended destinations and routes of all vehicles (V2I)

Conventional vs autonomous

vehicles

Conventional vehicles are driven by

people – we cannot predict their behavior (computational irreducibility of human's brain).

Microscopic models approximate drive in typical conditions.

It's difficult to predict human's response to a given atypical situation (assuming we don't have accurate driver's psychological model).

It is difficult to predict traffic state with sufficient accuracy when atypical situations occur.

Autonomous vehicles are driven by a

computer running (interactive) algorithm (computational model equivalent to Interactive Turing Machine).

Microscopic models could determine (or

approximate with good accuracy) drive in

typical conditions.

Machine's response to a given atypical situation could be computed (deterministic or stochastic algorithm).

Traffic state could be approximated with a good accuracy when atypical situations occur (in a long-term, sensitive dependence on initial conditions may make it impossible).

Implications

In the era of self-driving cars with V2X communication (which is expected to come...): O-D matrix will be given (almost) for free thanks to V2I communication.

Microscopic traffic simulation models might be calibrated online.

Traffic simulation could be run few orders of magnitude faster than realtime (High-Performance Computing clusters) and be very accurate.

Short-time prediction could be very accurate and fast thanks to simulations.

Current traffic state (positions, velocities, accelerations, routes of all cars) could be known by a traffic management center and updated frequently.

It may be possible to evaluate large number of configurations of traffic management parameters (e.g., configurations of traffic signals) and choose the best one –

Simulation-based traffic

management system

V2I

I2V

TMC

Large-scale traffic simulation

Details in my papers, e.g.,: P. Gora, 2009, „Traffic Simulation Framework

Initial results

Details are in my recent work (09.2015): „Application of

genetic algorithms and high-performance computing to

the Traffic Signal Setting problem”

Assuming we have a complete knowledge about current positions, speeds and routes of cars (and O-D matrix to simulate new cars, starting drive within the next 10 minutes):

● I was running simulations (using TSF) to assess quality of traffic signal settings (offsets) ● Genetic algorithm encoding offsets to genotypes (population: 400-900 genotypes, 10% best

genotypes selected for reproduction)

● Simulating 10 minutes of large-scale traffic (with over 100 000 cars) using microscopic models takes

several minutes, using mesoscopic model → a few seconds

● Genetic algorithm needs to assess many traffic signal settings to find good solutions, so:

– Mesoscopic model is better (faster, less accurate, but still can find acceptable solution)

– Parallelization of traffic signal settings evaluation is required (algorithm level, simulation level) – Other metaheuristics (e.g. Monte Carlo methods, Metropolis-Hastings, simulated annealing)

Initial results

Details are in my recent work (09.2015): „Application of genetic algorithms and

high-performance computing to the Traffic Signal Setting problem”

● Experiments conducted from January to July 2015, High-Performance Computing cluster

(University of Rzeszów):

– 40 computational nodes, 200 cores, computational power: 7.5 TeraFLOPS

– Up to 18.12% reduction of travel delay on the whole road network of Warsaw (~800 crossroad

with traffic signals)

– Parallelization on „Experiment level” (running many instances of GA in parallel, on many cores) – A single GA run: at least 10000 seconds ~= 3 hours → the methodology cannot be applied to

real-time traffic management yet, but...

– After introducing „Algorithm level parallelization” (evaluating TSS qualities in parallel within a

single GA run – is technically feasible) → ~100 seconds < 2 minutes (might be applied to real-time traffic management

– After introducing „Simulation level parallelization” (running a single traffic simulation in parallel –

potentially could be technically feasible):

● HPC cluster in Rzeszów → ~100 seconds (might be applied to real-time traffic

management)

Microscopic traffic simulation with

autonomous and connected cars

Inga Rüb, [email protected]

Idea: think of a single car as of an agent in a multiagent system

(details in our paper: „Traffic models for autonomous and connected

cars” - to be published in April 2016)

BDI concept:

Beliefs

Desires

Intentions

Actions:

slow down, speed up, change direction, complex actions...

knowledge of the environment (obstacles,

other cars and their planned routes)

general rules to follow and goals to obtain

plans for the nearest future (the next

simulation step)

Issues

The world is not ideal:

Issue Potential solution

Bugs in code Comprehensive tests Defects of crucial hardware parts (e.g.

sensors, communication devices, cars) Redundant functionalities, self-diagnostics, components self-and healing, self-repairing

Cryptographic attacks Secure communication protocols, communication limited to simple, verificable and indispensable commands Coexistence of autonomous/connected and

conventional vehicles Detecting other cars using sensors, new microscopic models. Pedestrians, cyclists Sensors scanning environment, detecting and predicting motion, communication with mobile devices, airbags

Bad weather conditions V2X / I2V communication to improve detection

Conclusions

●

Autonomous and connected cars may bring revolution in transportation and

traffic management (much less accidents, better traffic management). New

traffic models are required.

●

Large-scale traffic simulation might be applied to real-time traffic management

(e.g., (sub)optimal reaction to atypical conditions within a few seconds).

●

High-performance computing clusters (e.g., „PEGAZ” in Rzeszów, „Prometheus”

in Kraków) might be required (alternatives: GPU, quantum computer?).

●

This is just initial idea (and initial results). My goal was to show that this might be

technically feasible in the future, so some research efforts should be put on this

direction.

Thank you!!!

Questions?

➢

[email protected]

➢

http://www.mimuw.edu.pl/~pawelg

„Logic will get you from A to B. Imagination will take

you everywhere.” A. Einstein

„Sky is not the limit.”

The presentation was partially supported by the research grant 2011/01/D/ST6/06981 of the National Science Centre