Assuring City Scale Infrastructure Systems
Principal Investigators: Yair Amir, JHU Computer Science
and Tamim Sookoor, JHU Applied Physics Lab
Contributors
- JHU Computer Science, DSN Lab
- JHU Applied Physics Lab
Overview
AI systems are optimized for the average case. They cannot be used in critical systems that need to guarantee safety in worst case scenarios. The problem with AI systems is the long tail of edge cases that lead to failure situations. We want to gain the benefits of AI on the average case without incurring failures due to the long tail edge cases.
Proposed Approach
Designing a monitoring architecture that can take over from the AI with a safe controller when the AI system is at risk of breaching the system invariants. We intend to combine an invariant-based Black-Box Monitor with a White-Box Monitor that evaluates the confidence of the machine learning algorithm

Flagship Projects
Two ecosystem testbeds targeted at transportation and public safety domains
Presentations
Traffic Simulator Testbed Videos
Safe Controller
Average waiting time is 9.97 seconds
RL Model
50K training steps, average waiting time is 34.51 seconds
1M training steps, average waiting time is 14.52 seconds
3M training steps, average waiting time is 5.58 seconds
30M training steps, average waiting time is 0.29 second
Safe Controller with a Guarantee
Guarantee: vehicles that are going straight or turning right stop at most once at an intersection.
Average speed: 5.69047 (m/s)
A More Realistic Model
50K training steps, average speed is 1.97545 (m/s)
10M training steps, average speed is 3.01981 (m/s)
30M training steps, average speed is 5.66471 (m/s)
50.3M training steps, average speed is 6.00123 (m/s)
67.4M training steps, average speed is 6.48839 (m/s)
Switching Mechanism
Switch from AI Controller to Safe Controller:
switch when the average speed of past 50 seconds is below 4 (m/s)
Switch from Safe Controller to AI controller
Strategy 1:
After 60 seconds, switch to the AI controller when the average speed of the past 50 seconds is above 5 (m/s).
Strategy 2:
After 60 seconds, run a test simulation with the AI controller for 1000 steps. The test simulation uses an inflow rate of the past 50 seconds. If the average speed of the test simulation is above 5 (m/s), switch to the AI controller.
Results
Strategy 1: average speed (m/s) of all vehicles is 5.53297
Strategy 2: average speed (m/s) of all vehicles is 5.84685
Each of the following simulations has 9000 steps (900 seconds).
Between 1-300 seconds, the inflow is 500 vehicles per hour on each outside edge.
Between 301-600 seconds, the inflow is 700 vehicles per hour on one edge, and 100 vehicles per hour on other edges.
Between 601-900 seconds, the inflow becomes 500 vehicles per hour on each outside edge again.
Strategy 1
Strategy 2
Moving from grid to real world map