Regulating Highly Automated Robot Ecologies: Insights from Three User Studies

Transcription

1 Regulating Highly Automated Robot Ecologies: Insights from Three User Studies Wen Shen (UC, Irvine) Alanoud Al Khemeiri (Masdar Institute) Abdulla Almehrzi (Masdar Institute) Wael Al Enezi (Masdar Institute) Iyad Rahwan (MIT) Jacob W. Crandall (BYU)

2 Human Societies How do we achieve good human societies?

3 Human Societies How do we achieve good human societies? Strong central authority vs. strong individual rights

4 Societies of Robots? Self-driving cars Robotic buildings connected via a smart grid Financial Markets

5 Societies of Robots? Self-driving cars Robotic buildings connected via a smart grid Financial Markets Strong central authority vs. strong individual rights

6 Highly Automated Robot Ecologies Society of robots or systems Robots are independent owned by different stakeholders Robots are autonomous (from the perspective of the regulator)

7 Highly Automated Robot Ecologies Society of robots or systems Robots are independent owned by different stakeholders Robots are autonomous (from the perspective of the regulator) How can such systems be designed to produce good societal outcomes?

8 HARE are like what?

9 HARE are like what? Supervisory control systems

10 HARE are like what? Supervisory control systems

11 HARE are like what? Supervisory control systems Human Society

12 HARE are like what? Supervisory control systems Human Society

13 HARE are like what? Supervisory control systems Human Society Mechanism design problem

14 HARE are like what? Supervisory control systems Human Society Mechanism design problem

15 Challenge: Design efficient HARE 2 design parameters Regulatory power Robot autonomy (adaptability)

16 Example: Routing Game 60 B 75 A D C 300

17 Example: Routing Game 60 B 75 A D C 300 Vij = f (Nij, Cij)

18 Example: Routing Game 60 B 75 A D C 300 Vij = f (Nij, Cij) # of vehicles on link i-j

19 Example: Routing Game 60 B 75 A D C 300 Vij = f (Nij, Cij) # of vehicles on link i-j capacity of link i-j

20 Example: Routing Game 60 B 75 A D C 300 V ij / 1 1+e 0.25(N ij C ij ) +0.1

21 Example: Routing Game A C B D Regulator s Goal: Maximize throughput through node D 300 V ij / 1 1+e 0.25(N ij C ij ) +0.1

22 Example: Routing Game A C B D Regulator s Goal: Maximize throughput through node D 300 V ij / 1 1+e 0.25(N ij C ij ) +0.1 Needs to remove traffic congestion

23

24

25 Robot Behaviors

26 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) was the utility for arriving at dest

27 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) was the utility for arriving at dest Value of getting to node g

28 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) Value of getting to node g was the utility for arriving at dest Travel Cost

29 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) Value of getting to node g was the utility for arriving at dest Travel Cost Toll Cost

30 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) Value of getting to node g was the utility for arriving at dest Travel Cost Toll Cost Robot Autonomy (2 levels)

31 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) Value of getting to node g was the utility for arriving at dest Travel Cost Toll Cost Robot Autonomy (2 levels) Simple Estimate ct(i,g) assuming no congestion

32 Robot Behaviors u(i,g)=v(g) c t (i,g) c $ (i,g) Value of getting to node g was the utility for arriving at dest Travel Cost Toll Cost Robot Autonomy (2 levels) Simple Estimate ct(i,g) assuming no congestion Adaptive Estimate ct(i,g) using reinforcement learning

33 Regulatory Power Regulator s ability to change tolls

34 Regulatory Power Regulator s ability to change tolls 3 levels

35 Regulatory Power Regulator s ability to change tolls 3 levels None Regulator can do nothing

36 Regulatory Power Regulator s ability to change tolls 3 levels None Regulator can do nothing Limited Regulator can make limited toll changes

37 Regulatory Power Regulator s ability to change tolls 3 levels None Regulator can do nothing Limited Regulator can make limited toll changes Unlimited Regulator can make unlimited toll changes

38 Experimental Setup Regulatory Power Algorithmic Sophistication Simple Adaptive None Limited Unlimited Which one will be best?

39 Outcome Simple automation Adaptive automation % Optimal Throughput None Limited Unlimited Regulatory Power

40 Outcome Simple automation Adaptive automation % Optimal Throughput Algorithmic Sophistication Adaptive Regulatory Power None Limited Unlimited Simple 6 1 T2 T4 T4 T None Limited Unlimited Regulatory Power

41 Why Simple-Unlimited? Toll Adjustments ( per second) Simple automation Adaptive automation Algorithm Vehicle Type Node Preference Simple automation Adaptive automation 0.0 Limited Unlimited 0 Limited Unlimited Regulatory Power Regulatory Power (a) (b)

42 Why Simple-Unlimited? Toll Adjustments ( per second) Simple automation Adaptive automation Algorithm Vehicle Type Node Preference Simple automation Adaptive automation 0.0 Limited Unlimited 0 Limited Unlimited Regulatory Power Regulatory Power (a) (b) Given Unlimited Power, Regulators used power they didn t need

43 Why Simple-Unlimited? Toll Adjustments ( per second) Simple automation Adaptive automation Algorithm Vehicle Type Node Preference Simple automation Adaptive automation 0.0 Limited Unlimited 0 Limited Unlimited Regulatory Power Regulatory Power (a) (b) Given Unlimited Power, Regulators used power they didn t need

44 Why Simple-Unlimited? Toll Adjustments ( per second) Simple automation Adaptive automation Algorithm Vehicle Type Node Preference Simple automation Adaptive automation 0.0 Limited Unlimited 0 Limited Unlimited Regulatory Power Regulatory Power (a) (b) Given Unlimited Power, Regulators used power they didn t need Regulators had poorer models of robot behavior

45 Why Simple-Unlimited? Toll Adjustments ( per second) Simple automation Adaptive automation Algorithm Vehicle Type Node Preference Simple automation Adaptive automation 0.0 Limited Unlimited 0 Limited Unlimited Regulatory Power Regulatory Power (a) (b) Simple automation was easier to model

46 Automated Help B A D C

47 Automated Help Predict when the congestion will occur B A D C

48 Automated Help Predict when the congestion will occur Alert the regulator of predicted congestion B A D C

49 Automated Help Predict when the congestion will occur Alert the regulator of predicted congestion B A D C

50 Outcome Forecasting Yes No % Optimal Throughput Simple automation Adaptive automation Limited Unlimited Limited Unlimited Regulation

51 Outcome Forecasting Yes No % Optimal Throughput Simple automation Adaptive automation Limited Unlimited Limited Unlimited Regulation

52 Outcome Forecasting Yes No % Optimal Throughput Simple automation Adaptive automation Limited Unlimited Limited Unlimited Regulation Decision support made Simple-Limited worse!

53 Outcome Forecasting Yes No % Optimal Throughput Simple automation Adaptive automation Limited Unlimited Limited Unlimited Regulation Decision support made Simple-Limited worse! Why? Regulators had a poorer model of the cars.

54 Toward a General Theory 3 Forces : Adaptive robots -> Regulator must spend more time modeling Adaptive robots -> Regulators need more regulatory power More regulator power -> Decreased time modeling robots

55 Conclusions and Future Work

56 Conclusions and Future Work Data points that suggest less is more

57 Conclusions and Future Work Data points that suggest less is more Limited regulator power with simple robots produced the best results

58 Conclusions and Future Work Data points that suggest less is more Limited regulator power with simple robots produced the best results Just outliers? Or part of a general trend?

59 Conclusions and Future Work Data points that suggest less is more Limited regulator power with simple robots produced the best results Just outliers? Or part of a general trend? Can we find a way to do more with more?

60 Extras

61 Time Elapsed! Game length = 25 min Traffic Indicator Toll Adjustment Displayed after the! game was completed