Ant Colony Optimization for Air Traffic Conflict Resolution

Transcription

1 Ant Colony Optimization for Air Traffic Conflict Resolution Nicolas Durand, Jean-Marc Alliot DSNA/R&D/POM 1 July 1, DSNA/DTI R&D/Planing Optimization Modeling Team

2 Introduction Ant Colony Optimization Application to Conflict Resolution Conclusion

3 Conflict Resolution Current Situation : no effective tool for separating aircraft

4 Conflict Resolution Current Situation : no effective tool for separating aircraft New means : GPS capabilities (FMS enhancement), Data-Link communications Enhance Trajectory Prediction

5 Conflict Resolution Current Situation : no effective tool for separating aircraft New means : GPS capabilities (FMS enhancement), Data-Link communications Enhance Trajectory Prediction Pairwise conflicts Clusters

6 Conflict Resolution Current Situation : no effective tool for separating aircraft New means : GPS capabilities (FMS enhancement), Data-Link communications Enhance Trajectory Prediction Pairwise conflicts Clusters High complexity of the underlying problem

7 Conflict Resolution Current Situation : no effective tool for separating aircraft New means : GPS capabilities (FMS enhancement), Data-Link communications Enhance Trajectory Prediction Pairwise conflicts Clusters High complexity of the underlying problem Example : solving a n aircraft conflict in the horizontal plane n(n 1) 2 aircraft pairs 2 n(n 1) 2 connected components to explore

8 Conflict Resolution Current Situation : no effective tool for separating aircraft New means : GPS capabilities (FMS enhancement), Data-Link communications Enhance Trajectory Prediction Pairwise conflicts Clusters High complexity of the underlying problem Example : solving a n aircraft conflict in the horizontal plane n(n 1) 2 aircraft pairs 2 n(n 1) 2 connected components to explore Local optimization uneffective

9 Examples of existing algorithms Central & Global approaches Integer Linear programming

10 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming

11 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming Branch and Bound Intervals

12 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming Branch and Bound Intervals Genetic Algorithms

13 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming Branch and Bound Intervals Genetic Algorithms Autonomous approaches

14 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming Branch and Bound Intervals Genetic Algorithms Autonomous approaches Neural network

15 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming Branch and Bound Intervals Genetic Algorithms Autonomous approaches Neural network Repulsive forces

16 Examples of existing algorithms Central & Global approaches Integer Linear programming Semi-definite programming Branch and Bound Intervals Genetic Algorithms Autonomous approaches Neural network Repulsive forces Iterative approaches : give priorities to aircraft and use local optimization (for example : A algorithm).

17 ACO principles Use the environment as a medium of communication Home Food

18 ACO principles Use the environment as a medium of communication Mimic the ants trying to find the shortest path from their colony to food Home Food

19 ACO algorithm principle Ants deposite pheromones according to the quality of the path they find

20 ACO algorithm principle Ants deposite pheromones according to the quality of the path they find Ants more likely to follow paths with the most pheromones

21 ACO algorithm principle Ants deposite pheromones according to the quality of the path they find Ants more likely to follow paths with the most pheromones Add evaporation process to prevent algorithm from local convergence

22 ACO algorithm principle Ants deposite pheromones according to the quality of the path they find Ants more likely to follow paths with the most pheromones Add evaporation process to prevent algorithm from local convergence Stop when no more improvement

23 ACO for the Traveling Salesman Problem Ants sent on graph. Each ant builds complete path. Choice of next city influenced by pheromone quantity on paths. F O E A C D G B Vidéo

24 ACO for the Traveling Salesman Problem Ants sent on graph. Each ant builds complete path. Choice of next city influenced by pheromone quantity on paths. O A Ants deposite pheromones on the path chosen : τ ij (t) 1 Lij F E C D G B Vidéo

25 ACO for the Traveling Salesman Problem Ants sent on graph. Each ant builds complete path. Choice of next city influenced by pheromone quantity on paths. O A Ants deposite pheromones on the path chosen : τ ij (t) 1 Lij F E C D At each iteration, evaporate trails : τ ij ρ τ ij G B Vidéo

26 ACO for the Traveling Salesman Problem Ants sent on graph. Each ant builds complete path. Choice of next city influenced by pheromone quantity on paths. O A Ants deposite pheromones on the path chosen : τ ij (t) 1 Lij F E C D At each iteration, evaporate trails : τ ij ρ τ ij Stop when no more improvement G B Vidéo

27 n aircraft conflict example Conflict zone n aircraft

28 Maneuver modeling U V T1 W T0 Discretize time into timesteps 3 possible angles : 10, 20 or 30 degrees

29 Possible transitions U V W END U i+1 = U i V i+1 = V i + 6 W i+1 = V i U 1 = 1, V 1 = 6 and W 1 = 0 Number of possible states at timestep i= U i + V i + W i = 12 i 5

30 One ant per cluster or one ant per aircraft one ant one cluster one ant for n aircraft one ant per aircraft

31 One ant per cluster or one ant per aircraft one ant one cluster for n aircraft and t timesteps : (12 t 5) n trails. one ant for n aircraft one ant per aircraft

32 One ant per cluster or one ant per aircraft one ant one cluster for n aircraft and t timesteps : (12 t 5) n trails. For n = 5 and t = 10 : more than trails one ant for n aircraft one ant per aircraft

33 One ant per cluster or one ant per aircraft one ant one cluster for n aircraft and t timesteps : (12 t 5) n trails. For n = 5 and t = 10 : more than trails one ant one aircraft one ant for n aircraft one ant per aircraft

34 One ant per cluster or one ant per aircraft one ant one cluster for n aircraft and t timesteps : (12 t 5) n trails. For n = 5 and t = 10 : more than trails one ant one aircraft for n aircraft and t timesteps : n (12 t 5) trails. one ant for n aircraft one ant per aircraft

35 One ant per cluster or one ant per aircraft one ant one cluster for n aircraft and t timesteps : (12 t 5) n trails. For n = 5 and t = 10 : more than trails one ant one aircraft for n aircraft and t timesteps : n (12 t 5) trails. For n = 30 and t = 20 : more than 7050 trails instead of one ant per aircraft one ant for n aircraft

36 Initial amount of pheromones on the graph END 1 1

37 Algorithm description (1) Each path is given a score (the smaller, the better)

38 Algorithm description (1) Each path is given a score (the smaller, the better) U = +0, V = +2 and W = +1

39 Algorithm description (1) Each path is given a score (the smaller, the better) U = +0, V = +2 and W = +1 Conflict no pheromones

40 Algorithm description (1) Each path is given a score (the smaller, the better) U = +0, V = +2 and W = +1 Conflict no pheromones No conflict τ = n n out n τ 0 s path where n out is the number of lost ants, τ 0 the original quantity of pheromones, and s path the score of the path followed by the ant.

41 Algorithm description (1) Each path is given a score (the smaller, the better) U = +0, V = +2 and W = +1 Conflict no pheromones No conflict τ = n n out n τ 0 s path where n out is the number of lost ants, τ 0 the original quantity of pheromones, and s path the score of the path followed by the ant. At each node, the next edge is chosen with a probability depending on its quantity of pheromones.

42 Algorithm description (2) An evaporation principle : the amount of pheromones is decreased by x% (in the examples x = 10%) at the end of each iteration.

43 Algorithm description (2) An evaporation principle : the amount of pheromones is decreased by x% (in the examples x = 10%) at the end of each iteration. Ending criteria : the score obtained by each bunch of ants no longer decreases.

44 Example of 5 aircraft conflict resolution 18 iterations - score=89

47 Algorithm improvement : constraint relaxation High density areas no ants are able to solve every conflict

48 Algorithm improvement : constraint relaxation High density areas no ants are able to solve every conflict Relax the conflict resolution constraint : accept ants with r remaining conflicts

49 Algorithm improvement : constraint relaxation High density areas no ants are able to solve every conflict Relax the conflict resolution constraint : accept ants with r remaining conflicts When solutions are found for a certain number of ants, the constraint is reinforced

50 Algorithm improvement : constraint relaxation High density areas no ants are able to solve every conflict Relax the conflict resolution constraint : accept ants with r remaining conflicts When solutions are found for a certain number of ants, the constraint is reinforced Example : define r as the minimum number of conflicts of the least conflicting ant

51 Algorithm improvement : constraint relaxation High density areas no ants are able to solve every conflict Relax the conflict resolution constraint : accept ants with r remaining conflicts When solutions are found for a certain number of ants, the constraint is reinforced Example : define r as the minimum number of conflicts of the least conflicting ant r is the number of allowed conflicts per ant at the first generation.

52 Algorithm improvement : constraint relaxation High density areas no ants are able to solve every conflict Relax the conflict resolution constraint : accept ants with r remaining conflicts When solutions are found for a certain number of ants, the constraint is reinforced Example : define r as the minimum number of conflicts of the least conflicting ant r is the number of allowed conflicts per ant at the first generation. Reduce r when the number of ants having less than r conflicts is higher than n r

53 Example of 30 aircraft conflict resolution generation: 0-4 conflicts max - 9 aircraft

57 Example of 30 aircraft conflict resolution generation: 45-1 conflict max - 20 aircraft

60 Vidéo Example of 30 aircraft conflict resolution generation: 65-0 conflict max - 30 aircraft

61 Conclusion The modeling can be extended to any trajectory

62 Conclusion The modeling can be extended to any trajectory Complexity depends on the number of alternate paths available

63 Conclusion The modeling can be extended to any trajectory Complexity depends on the number of alternate paths available Results will be compared to the existing ERCOS (using GAs)

64 Conclusion The modeling can be extended to any trajectory Complexity depends on the number of alternate paths available Results will be compared to the existing ERCOS (using GAs) Stochastic optimization : no guarantee of solution or optimum

65 Conclusion The modeling can be extended to any trajectory Complexity depends on the number of alternate paths available Results will be compared to the existing ERCOS (using GAs) Stochastic optimization : no guarantee of solution or optimum No effective tool offered to controllers without significant enhancement of ground TP