Master Degree in Computer Engineering Thesis Slides: Road Traffic Rules Synthesis Using Grammatical Evolution

1. Road Traffic Rules Synthesis
Using Grammatical Evolution
07 June 2017
University of Trieste
Supervisor: Eric Medvet
Co-supervisors: Alberto Bartoli
Lorenzo Castelli
Author: Jacopo Talamini

2. Road Traffic (Present Scenario)
Road traffic agents:
 Human drivers
Road traffic rules:
 Traffic rules designed for humans
2

3. Road Traffic (Future Scenario)
Road traffic agents:
 Human drivers
 Driverless cars
Road traffic rules:
 Traffic rules designed for humans
Do we need other rules or are the same ones suitable for both agents?
3

4. Road Traffic Agents (I)
 Human drivers
 Driverless cars
Same purposes:
 Reach a destination
 Avoid accidents
 Comply to the rules
4

5. Road Traffic Agents (II)
 Human drivers
 Driverless cars
Different nature with respect to:
 Reflexiveness
 Reliability
 Making decisions
 …
5

6. Re-writing the rules (I)
If we could re-write the rules from scratch, we would consider:
 Traffic efficency (reach destination in lowest time possible)
 Traffic safety (avoiding accidents)
How to write those rules?
6

7. Re-writing the rules (II)
If we could re-write the rules from scratch, we would consider:
 Traffic efficency (reach destination in lowest time possible)
 Traffic safety (avoiding accidents)
How to write those rules?
Proposed solution:
consider an evolutionary approch to write the rules automatically
7

8. Proposed Solution (I)
Evolutionary approach:
 Individual := set of rules
 Fitness := (efficiency, safety), simulated with those rules
8

9. Proposed Solution (II)
Evolutionary approach:
 Individual := set of rules
 Fitness := (efficiency, safety), simulated with those rules
Contributions made:
1. Road Traffic model
2. Language for the rules
3. GE-based rules generation
4. Experimental evaluation
9

10. Road Traffic model (I)
Trade-off:
 Detailed enough to resemble real traffic
 Simple enough to allow for simulation
Describes:
 The physical world
 The driver’s behaviour
10

11. Road Traffic model (II)
Infrastructure:
 Road graph
(include concepts like road segments, crossroads, lanes in sections)
Agents:
 Move according to the driver’s algorithm
 Continuous position in the sense of the road sections length
 Discrete position on the lanes
 Speed
 Status (collisions)
 …
11

12. Road Traffic model (III)
12

13. Driver’s Algorithm (I)
Each car moves according to the driver’s algorithm
Input:
 Input variables (info about closest cars)
 State variables (distance of view, …)
Output:
 a list A of actions:
Examples: accelerate, move on the left lane, ..
13

14. Driver’s Algorithm (II)
14

15. Driver’s Algorithm (III)
15

16. Driver’s Algorithm (IV)
Goals:
 No route to travel, just cross a target distance
 Possibly travel at maximum speed
 Avoid hitting other cars
 Perform always the desired action
16

17. Driver vs Rules-aware driver (I)
At each time step each driver performs exactly one action from A
Driver’s performed action:
 The first action on the list (most preferred action)
17

18. Driver vs Rules-aware driver (II)
At each time step each driver performs exactly one action from A
Driver’s performed action:
 The first action on the list (most preferred action)
Rules-aware driver’s performed action:
 The first action a in A which breaks the lowest number of rules
18

19. Driver vs Rules-aware driver (III)
At each time step each driver performs exactly one action from A
Driver’s performed action:
 The first action on the list (most preferred action)
Rules-aware driver’s performed action:
 The first action a in A which breaks the lowest number of rules
Why driver and rules-aware driver?
19

20. Driver vs Rules-aware driver (IV)
Why driver and rules-aware driver?
 Driver: free to drive as desired, no constraints
 Rules-aware driver: among preferred actions chooses a permitted one
20

21. Rules-aware driver prediction algorithm
At each time step rules-aware driver perform a prediction:
 For each action in A
 For k time steps in the future
 Assuming the same action for each time step
Result:
Action that will break the lowest number of rules
The lowest number of rules might be > 0 !
21

22. Language for the rules (I)
Rule:
 A predicate that can be evaluated to true or false
 Rules-aware drivers consider rules to choose the performed action
 All agents consider the same rules
Works on:
 Car status (speed, position,..)
 Other cars relative distances and speeds
 Current road section/intersection
22

23. Language for the rules (II)
Rules are expressed as the context-free grammar:
23

24. Language for the rules (III)
Rules are expressed as the context-free grammar:
24Goal: express rules in a human-readable way

25. Examples of rules (I)
Examples of possible rules:
 «stay on the rightmost free lane»
 «mantain a low speed near an intersection»
 «keep a minimum distance from the car ahead»
25

26. Example of rules (II)
Examples of possible rules:
 «stay on the rightmost free lane»
 «mantain a low speed near an intersection»
 «keep a minimum distance from the car ahead»
26

27. Hand-written rules
 Design inspired by real traffic rules
 Provide a first example of rules to the model
 Compared then with rules generated with GE
27

28. Evolutionary Computation (I)
We know :
 the nature of the solution
 A way to evaluate a candidate solution
Basic scheme:
 Dynamic population of individuals (individual die and are born)
 Fittest individuals survive and reproduce more than others
 Offspring inherit some character from the parents
28

29. Evolutionary Computation (II)
 Generational model
At each time tick we have:
 A population of parents
 A population of offspring (built from parents)
29

30. Evolutionary Computation (III)
Individual:
 Phenotype := candidate solution for the considered problem
 Genotype := its inner representation used for manipulations
Given the genotype of an individual, it is possible to obtain:
 Its phenotype
 Its fitness
30

31. Grammatical Evolution
Individual:
 Phenotype := string of a Language
 Genotype := binary string
Given the genotype of an individual, it is possible to obtain:
 Its phenotype: through mapping function (genotype -> phenotype)
 Its fitness: the goodness of the phenotype obtained
31

32. GE-based rules generation
In the considered scenrario:
 Solution space as Context-Free Grammar
 Individual := string of the defined Context-Free Grammar
 Fitness := linear combination of (traffic efficiency, traffic safety)
32

33. Fitness (I)
 Traffic efficiency := average speed ratio (ASR)
 (1 – AsR) =
 Traffic safety := 1 – collision-per-time (CpT)
 CpT =
Minimize the linear combination (1 – AsR, CpT):
33

34. Fitness (II)
Meaning of nsim:
The vaule is computed over many (nsim) simulations
It allows to detect common trends in different simulations
Minimize the linear combination (1 – AsR, CpT):
34

35. Experimental Evaluation (I)
Traffic model:
 Is it sound?
 Find values for parameters
GE-based rules generation:
 Are generated rules better than no rules?
 Are generated rules better than hand-written rules?
35

36. Experimental Evaluation (II)
Traffic model:
 Is it sound?
 Find values for parameters
GE-based rules generation:
 Are generated rules better than no rules?
 Are generated rules better than hand-written rules?
For both: how does injected traffic affect efficiency and safety
36

37. Traffic model validation (I)
5 road sections, up to 20 cars, values averaged on 10 simulations
37

38. Traffic model validation (II)
5 road sections, up to 20 cars, values averaged on 10 simulations
38

39. Traffic model validation (III)
5 road sections, up to 20 cars, values averaged on 10 simulations
39

40. Traffic model validation (IV)
30 runs, each with 100 individuals, 100 generations, ncar fixed!
 Slightly higher efficiency
 Higher safety 40

41. Congestion scenario (I)
41
Congestion:
 More traffic injected does not increase overall distance

42. Congestion scenario (II)
42
Congestion:
 More traffic injected does not increase overall distance
 Consider the case of hand-written rules

43. Congestion scenario (III)
43

44. Congestion scenario (IV)
44

45. Congestion scenario (V)
Results:
With high traffic injected (congestion)
GE vs. hand-written rules:
 Longer overall distance
 Less overall collisions
45

46. Conclusions
Future work:
 More fine-grained model
 Extend the model to other types of traffic
 …
46

47. Rules obtained with GE

48. Hand written rules