Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational...
Transcript of Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational...
![Page 1: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/1.jpg)
Cathy Wu | UC Berkeley
Future of AI in Transportation
TRB Executive Committee, Policy Session on AI in Transportation, 2020
Assistant Professor
LIDS, CEE, IDSS
![Page 2: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/2.jpg)
Outline
• Part 1: Why AI and why now?
• Part 2: What is AI good for?
• Part 3: Why AI instead of other methods?
![Page 3: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/3.jpg)
Outline
• Part 1: Why AI and why now?
• Part 2: What is AI good for?
• Part 3: Why AI instead of other methods?
![Page 4: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/4.jpg)
Complexities compound transportation issues today
• Connectivity and automation (smartphones, CAVs, 5G)
• Technology-enabled business models (TNCs, routing apps, micromobility)
• Climate change (extreme weather, supply chain disruptions)
• International affairs (trade disputes)
![Page 5: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/5.jpg)
Critical Issues in Transportation 2019
1. Transformational Technologies and Services: Steering the Technology Revolution
2. Serving a Growing and Shifting Population3. Energy and Sustainability: Protecting the Planet4. Resilience and Security: Preparing for Threats5. Safety and Public Health: Safeguarding the Public6. Equity: Serving the Disadvantaged7. Governance: Managing Our Systems8. System Performance and Management: Improving
the Performance of Transportation Networks9. Funding and Finance: Paying the Tab10. Goods Movement: Moving Freight11. Institutional and Workforce Capacity: Providing a
Capable and Diverse Workforce12. Research and Innovation: Preparing for the Future
9 of 12 critical issues exacerbated by growing complexity of transportation systems.
![Page 6: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/6.jpg)
Why AI and why now?
• Change is outpacing existing methodology for reliable transportation systems.
• Opportunity in data: all this complexity is increasingly captured (sensors, smartphones).
• Strength of AI, especially modern deep learning: extracting useful information out of a sea of data.
AI can help overcome complexity.
Image credit: PCMag
![Page 7: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/7.jpg)
Outline
• Part 1: Why AI and why now?
• Part 2: What is AI good for?
• Part 3: Why AI instead of other methods?
![Page 8: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/8.jpg)
Challenge: safety
Application
• Pedestrian safety systems
• In-vehicle safety systems
• Failures of infrastructure, vehicles, equipment
AI solution
• Predict potential accidents
• Context-aware technology
• Prediction of failures, automated inspections
AI to leverage the complexity to identify unsafe events
![Page 9: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/9.jpg)
Challenge: congestion
Application• Synchronized modalities
(MoD, bus, train, subway, bike)
• MoD curb-side management• Demand and mode shift• Advanced load balancing,
scheduling, and vehicle right-sizing based on preferences
• AVs for traffic smoothing
AI solution• Demand prediction
• Activity recommendation• Personalization and
preference inference
• Automatically learn vehicle controllers
AI to manage the complexity and coordinate supply and demand
![Page 10: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/10.jpg)
Emerging and cross-cutting
Application• Impact assessment
(new modes, regulations and pricing schemes, business models)
• Coordination among city functions (transportation, maintenance/works, energy, water, waste)
• Freight + AI• Immobility solutions
(virtual presence, augmented/virtual reality (AR/VR), telecommuting, co-working)
AI solution• Learned recommendations of
rules and regulations• Holistic prediction of city
demands
• Predict what people want to buy• Synthetic avatars
AI to transcend complexity and evolve the transportation system
![Page 11: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/11.jpg)
Outline
• Part 1: Why AI and why now?
• Part 2: What is AI good for?
• Part 3: Why AI instead of other methods?
![Page 12: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/12.jpg)
CloudCAVs for congestion mitigation
What is the potential impact on traffic congestion of automating a fraction of vehicles?
Focus: impact of vehicle kinematics
![Page 13: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/13.jpg)
Traffic jams1955
Partial differential equations (PDE)
Setting: 22 human drivers
Instructions: drive at 19 mph.
No traffic lights, stop signs, lane changes.
900 papers on PDEs for traffic
Sugiyama, et al.
2008 2019
![Page 14: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/14.jpg)
Video credits: NewScientist.com
Traffic jams1955
Partial differential equations (PDE)
Setting: 22 human drivers
Instructions: drive at 19 mph.
No traffic lights, stop signs, lane changes.
Traffic jams still form.
900 papers on PDEs for traffic
Sugiyama, et al.
2008 2019
![Page 15: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/15.jpg)
Mixed autonomy traffic1955
Partial differential equations (PDE)
900 papers on PDEs for traffic
Sugiyama, et al.
2008 2018
Automated
Observed
Unobserved
Setting: 22 human drivers
Instructions: drive at 19 mph.
No traffic lights, stop signs, lane changes.
Traffic jams still form.
![Page 16: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/16.jpg)
Decisions in urban systems:Vehicle accelerations
Tactical maneuvers
Transit schedules
Traffic lights
Land use
Parking
Tolling
…
Deep reinforcement learning (RL)
Agent
Environment
action at
state st
reward rt
rt+1
st+1
Global rewardsAverage velocity
Energy consumption
Travel time
Safety, comfort
Goal: learn policy (decision rule) to maximize long-term reward
DQN (2015) TRPO (2015) AlphaGo (2016)
Deep RL: methods to optimize a policy (a deep neural network) to maximize long-term reward in complex sequential decision problems.
![Page 17: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/17.jpg)
Setting: 1 AV, 21 human
Experiment• Goal: maximize average velocity• Observation: relative vel and headway
• Action: acceleration• Policy: multi-layer perceptron (MLP)• Learning algorithm: policy gradient
Results• 1 AV: +49% average velocity• First near-optimal controller for single-lane• Uniform flow at near-optimal velocity• Generalizes to out-of-distribution densities
Wu, et al. CoRL, 2017; Wu, et al. IEEE T-RO, in review.
AV offAV on
Mixed autonomy traffic (AI solution)
1955
Wu, et al.
20192008
2017
Automated
Observed
Unobserved
Sugiyama, et al.
![Page 18: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/18.jpg)
1955
Stern, et al.
20192008
2017
Sugiyama, et al.
Instructions: follow the vehicle in front, and close gaps. No tail-gaiting!
AV: Hand-tuned model-based controller (PI saturation)
Traffic jams diminished. 1 AV: +14% average velocity (vs. 49%)
Mixed autonomy traffic (non-AI solution)
![Page 19: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/19.jpg)
Sugiyama, et al.
On/off-rampSingle-lane
Grid networkBottleneck
Intersection
Straight highwaySignalized intersection
Multi-lane
AI + Traffic LEGO blocks Benchmarks for autonomy in transportation
Wu, et al. IEEE ITSC, 2017; Wu, et al. IEEE T-RO, in review; Vinitsky, Kreidieh, …, Wu, et al. CoRL, 2018.
5-10% AVs
+142%+30%+49%
+25%
+60%
1955
Wu, et al.
20172008
2019
![Page 20: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/20.jpg)
Setting: No AVs, 100% IDM
Setting: 10% AVs, 90% IDM
1480 veh/hrPhenomenon: capacity drop
1800 veh/hr
RL + increasing complexity (current work)
Results:• 25% improvement• Avoids capacity drop• Learned policy
transfers to different inflow rates, number of lanes, and percent of autonomous vehicles
Successful transfer:Network: 8 > 4 > 2 Bottleneck
Network: 8 > 4 > 2 > 1 Bottleneck
Scenario: varying inflow rates, varying % AVs.
Zhongxia Yan
![Page 21: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/21.jpg)
Where we are, where next?Full-scale regional network
SanFranciscoBayBridgeSanFranciscoDowntown
• Key idea: AI has the potential to keep pace with increase in complexity.
• Research challenge: Is there a limit for the level of complexity that AI can handle?
• Relies on access to data, which is increasingly privatized.
Physical deployment
Insights for transportation planning
![Page 22: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/22.jpg)
Collaborators & Partners
Eugene VinitskyBerkeley
Aboudy KreidiehBerkeley
Cathy Wu | UC Berkeley
Kanaad ParvateBerkeley
Kathy JangBerkeley
Nishant KheterpalBerkeley
Ananth KuchibhotlaBerkeley
Leah DicksteinBerkeley
Nathan MandiBerkeley
Alexandre BayenBerkeley
Zhongxia YanMIT
![Page 23: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/23.jpg)
New lab!
Zhongxia Yan PhD student
Weizi Li Postdoc
77 Massachusetts Avenue, 32-D608 T 617 253 2142Cambridge, MA 02139-4307 U.S.A. F 617 253 3578MASSACHUSETTS INSTITUTE OF TECHNOLOGY
77 Massachusetts Avenue, 32-D608 T 617 253 2142Cambridge, MA 02139-4307 U.S.A. F 617 253 3578MASSACHUSETTS INSTITUTE OF TECHNOLOGY
Vindula Jayawardana PhD student
![Page 24: Future of AI in Transportationonlinepubs.trb.org/onlinepubs/excomm/20-01-Wu.pdfTransformational Technologies and Services: Steering the Technology Revolution 2. Serving a Growing and](https://reader033.fdocuments.us/reader033/viewer/2022051812/602e47f89d656d4ce718b2d3/html5/thumbnails/24.jpg)
Future of AI in transportationMain message:
• Critical issues in transportation are exacerbated by growing complexity and increasing pace of change in the world.
• There is an opportunity to overcome these challenges by developing AI to leverage, manage, and transcend complexity and evolve our transportation systems.
Takeaways:
• Strong potential for AI in future solutions in safety, congestion, and emerging and cross-cutting applications.
• AI-based solutions may have a chance at keeping up with the pace of change in the world. Requires further investigation.
• AI-based solutions rely on access to increasingly privatized data.