1. Distributed Reinforcement Learning for a Traffic Engineering Application Mark D. Pendrith DaimlerChrysler Research & Technology Center Presented by: Christina Schweikert

2. Distributed Reinforcement Learning for Traffic Engineering Problem  Intelligent Cruise Control System  Lane change advisory system based on traffic patterns  Optimize a group policy by maximizing freeway utilization as shared resource  Introduce 2 new algorithms (Monte Carlo-based Piecewise Policy Iteration, Multi-Agent Distributed Q-learning) and compare their performance in this domain

3. Distronic Adaptive Cruise Control

4.  Signals from radar sensor, which scans the full width of a three-lane motorway over a distance of approximately 100m and recognizes any moving vehicles ahead  Reflection of the radar impulses and the change in their frequency enables the system to calculate the correct distance and the relative speed between the vehicles

5. Distronic Adaptive Cruise Control  Distance to vehicle in front reduces - cruise control system immediately reduces acceleration or, if necessary, applies the brake  Distance increases – acts as conventional cruise control system and, at speeds of between 30 and 180 km/h, will maintain the desired speed as programmed  Driver is alerted of emergencies

6. Distronic Adaptive Cruise Control  Automatically maintains a constant distance to the vehicle in front of it, prevent rear-end collisions  Reaction time of drivers using Distronic is up to 40 per cent faster than that of those without this assistance system

7. Distributed Reinforcement Learning  State – agents within sensing range  Agents share a partially observable environment  Goal - Integrate agents’ experiences to learn an observation-based policy that maximizes group performance  Agents share a common policy, giving a homogeneous population of agents

8. Traffic Engineering Problem  Population of cars, each with a desired traveling speed, sharing a freeway network  Subpopulation with radar capability to detect relative speeds and distances of cars immediately ahead, behind, and around them

9. Problem Formulation  Optimize average per time-step reward, by minimizing the per-car average loss at each time step v d (i) desired speed of car i v a (i) actual speed of car i n number of cars in simulation at time-step

10. State Representation  View of the world for each car represented by 8-d feature vector – relative distances and speeds of surrounding cars ALACAR CLCarCR BLBCBR

11. Pattern of Cars in Front of Agent ALACAR  0 – lane is clear (no car in radar range or nearest car is faster than agent’s desired speed)  1 – fastest car less than desired speed  2 – slower  3 - still slower

12. Pattern of Cars Behind Agent ALACAR  0 – lane is clear (no car in radar range or nearest car is slower than agent’s current speed)  1 – slowest car faster than desired speed  2 – faster  3 - still faster

13. Lane Change CLCARCR  0 – lane change not valid  1 – lane change valid If there is not a safe gap in front and behind, land change is illegal.

14. Monte Carlo-based Piecewise Policy Iteration  Performs approximate piecewise policy iteration where possible policy changes for each state are evaluated by Monte Carlo estimation  Piecewise - Policy for each state is changed one at a time, rather than in parallel  Searches the space of deterministic policies directly without representing the value function

15. Policy Iteration  Start with arbitrary deterministic policy for given MDP  Generate better policy by calculating best single improvement in policy possible for each state (MC)  Combine all changes to generate successor policy  Continue until no improvement is possible – optimal policy

16. Multi-Agent Distributed Q-Learning Q-Learning  Q-value estimates updated after each time step based on state transition after action is selected  For each time step, only one state transition and one action used to update Q-value estimates  In DQL, there can be as many state transitions per time step as there are agents

17. Multi-Agent Distributed Q-Learning  Takes the average backup value for a state/action pair over all agents that selected action a from state s at the last time step  Q max component of backup value is calculated over actions valid for a particular agent to select at the next time-step

18. Simulation for Offline Learning Advantages: o Since true state of the environment is known, can directly measure loss metric o Can be run faster, many long learning trials o Safety Learn policies offline then integrate into intelligent cruise control system with lane advisory, route planning, etc.

19. Traffic Simulation Specifications  Circular 3 lane freeway 13.3 miles long with 200 cars  Half follow “selfish drone” policy  Rest follow current learnt policy and active exploration decisions  Gaussian distribution of desired speeds, mean of 60 mph  Cars have low level collision avoidance, differ in lane change strategy

20. Experimental Results  Selfish drone policy – consistent per- step reward of -11.9 (each agent traveling 11.9 below desired speed)  APPIA and DQL found policies 3-5% better  Best policies with “look ahead” only  “look behind” model provided more stable learning  “look behind” outperforms “look ahead” at times when good policy is lost