Project: Multi-Modal Optimization In Surtrac

Keywords: ML algorithms, generalized predictive models, scientific computing, resource allocation, distributed simulations

Goal: Active management of transit vehicles (e.g., buses) in Surtrac

Problems:

  • Transit vehicle flow patterns differ significantly from regular cars
  • This is due to frequent stops and uncertain dwell times
  • This causes significant traffic flow disruptions, increased wait & delay times

Solutions:

  • Develop ML-predictive algorithms to accurately estimate dwell times
  • Extend the flow model in Surtrac to incorporate this information
  • Weight clusters by mode type to give priority
surtrac-multi-modal

Stage-I: Analyze Statistical Trends In DT

Data

  • Data – 2+ years of Port Authority data on buses moving along the Centre Avenue corridor (> 3M data points)
  • For each data point: stop id, day of week, direction, arrival times, departure times, miles (minutes) from last stop, load, number of passengers on (off), etc.

Findings

  • There is strong seasonal correspondence of dwell times from year to year.
  • Dwell times vary considerably by stop.
  • Dwell times also vary considerably within peak periods.
Dwell-time-variance-by-stop

Stage II: Generalized Predictive DT Models

Stage Overview

Goal:

  • Probabilistic models of bus dwell times for each bus route and bus stop
  • Highly stochastic behavior requires a technique that needs only a few data samples

Approach: Hierarchical, Bayes nets

Performance

  • Accurate predictions after just 20 data samples
  • Significantly outperforms standard linear regression techniques, online linear regression techniques, and deep learning
Stage-II- Generalized Predictive DT Models

Stage-III: Improve Algorithmic Performance

Stage Overview

Goal:

  • Improve the computational scalability of the dwell time prediction model
  • The model needs to numerically approximate an 8-10 dimensional integral
  • Computation needs ~ 7–250 seconds
  • Existing model can’t be used in real-time

Approach:

  • Identified a numerical integration algorithm used to analyze astronomy data
  • Extended the approximation algorithm to real-time computation
  • Incorporated it into a Bayesian sequential learning framework

Performance:

  • Average computational time reduced to 0.23 seconds without compromising accuracy
  • Significantly outperforms other numerical integration algorithms (Metropolis-Hastings, Hamiltonian Monte Carlo)
bayesian-sequential-learning
nested-sampling
LSTM
ARIMA

Application: Grant Street Area Simulation

 

Existing Timing Plans

 

surtrac

Get in Touch