almost optimal strategies in priced timed games
play

Almost-Optimal Strategies in Priced Timed Games Patricia Bouyer 1 , - PowerPoint PPT Presentation

Almost-Optimal Strategies in Priced Timed Games Patricia Bouyer 1 , Kim G. Larsen 2 , Nicolas Markey 1 , and Jacob Illum Rasmussen 2 1 Lab. Specification et Verification, ENS Cachan & CNRS, France 2 Dept of Computer Science, Aalborg University,


  1. Memorylessness and optimality Fact In our PTGAs, optimal strategies do not always exist. Fact When optimal strategies exist, they might require some memory. Example x =1 � x ≤ 1 p =2 ˙ x < 1 , x :=0 x > 0 p =1 ˙ An optimal strategy depends on the date at which the blue state is entered. But there is a memoryless ε -optimal strategy.

  2. Decidability of 1PTGAs Definition Given ε > 0 and N ∈ Z + , a strategy σ is ( ε, N ) acceptable if σ is ε -optimal and memoryless, there is a partition ( I n ) n ≤ N of [0 , M ] (where M is the maximal constant of the guards and invariants of the game) s.t., for any q ∈ Q c , x �→ σ ( q , x ) is constant on each I n .

  3. Decidability of 1PTGAs Definition Given ε > 0 and N ∈ Z + , a strategy σ is ( ε, N ) acceptable if σ is ε -optimal and memoryless, there is a partition ( I n ) n ≤ N of [0 , M ] (where M is the maximal constant of the guards and invariants of the game) s.t., for any q ∈ Q c , x �→ σ ( q , x ) is constant on each I n . Main Theorem For every location, the optimal cost is computable and is piecewise affine. There exists N ∈ Z + s.t., for any ε > 0, we can effectively compute an ( ε, N )-acceptable (thus, almost-optimal and memoryless) strategy.

  4. Simplifying the problem We restrict to TGAs with maximal constant 1 (in clock constraints)

  5. Simplifying the problem We restrict to TGAs with maximal constant 1 (in clock constraints) Example x =1 x =1 x =1 x :=0 x :=0 x :=0 x < 3 x < 1 x =1 x =1 x =1 p =2 ˙ p =2 ˙ p =2 ˙ p =2 ˙ p =2 ˙ x ≤ 4 x ≤ 1 x ≤ 1 x ≤ 1 x ≤ 1 x :=0 x :=0 x :=0 x ≥ 2 x =1 x =1 x =1 x :=0 x :=0 x :=0

  6. Simplifying the problem We restrict to strongly-connected TGAs without resets.

  7. Simplifying the problem We restrict to strongly-connected TGAs without resets. Example x ≤ 1 p =4 ˙ p =1 ˙ p =3 ˙ p =1 ˙ x :=0 x :=0

  8. Simplifying the problem We restrict to strongly-connected TGAs without resets. Example x ≤ 1 p =4 ˙ p =1 ˙ p =3 ˙ p =1 ˙ x :=0 x :=0

  9. Simplifying the problem We restrict to strongly-connected TGAs without resets. Example x ≤ 1 p =4 ˙ p =1 ˙ x ≤ 1 p =4 ˙ p =1 ˙ p =3 ˙ p =1 ˙ x :=0 x :=0 p =3 ˙ p =1 ˙ x :=0 x :=0

  10. Simplifying the problem We restrict to strongly-connected TGAs without resets. Example x ≤ 1 p =4 ˙ p =1 ˙ x ≤ 1 p =4 ˙ p =1 ˙ p =3 ˙ p =1 ˙ x :=0 p =3 ˙ p =1 ˙ x :=0 + ∞ x :=0

  11. Simplifying the problem We restrict to strongly-connected TGAs without resets. m x :=0 n G m 1 G ′ + ∞ x :=0 n 2

  12. Simplifying the problem We restrict to strongly-connected TGAs without resets. Theorem m x :=0 OptCost G ( q , x ) = OptCost G ′ ( q 1 , x ) . n G m 1 G ′ + ∞ x :=0 n 2

  13. Simplifying the problem We restrict to strongly-connected TGAs without resets. Theorem m x :=0 OptCost G ( q , x ) = OptCost G ′ ( q 1 , x ) . n G Theorem If σ ′ is ( ε ′ , N ′ )-acceptable in G ′ , then m 1  σ ′ ( q 2 , x ) G ′   σ ( q , x ) = if Cost( q 2 , x ) ≤ Cost( q 1 , x )  σ ′ ( q 1 , x ) otherwise  + ∞ x :=0 n 2 is (2 ε ′ , N ′ )-acceptable in G .

  14. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  15. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  16. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  17. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  18. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  19. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  20. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  21. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 x :=0 no resetting transitions.

  22. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 no resetting transitions.

  23. Simplifying the problem Reduced to x :=0 strongly-connected PTGAs clock is bounded by 1 no resetting transitions.

  24. Main theorem with outside cost-functions Theorem Let G be a strongly-connected non- resetting 1PTGA with outside cost- functions. p =3 ˙ OptCost G is computable; p =1 ˙ in each location, function x ≤ 1 x �→ OptCost G ( q , x ) p =5 ˙ p =1 ˙ is decreasing, piecewise affine and continuous. Its finitely many segments either have slope − c where c is the price of some locations, or are fragments of the outside cost-functions; There exists N ∈ Z + s.t., for any ε > 0 , we can compute an ( ε, N ) -acceptable strategy σ .

  25. Operations on cost functions: controllable locations p = 3 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  26. Operations on cost functions: controllable locations p = 3 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  27. Operations on cost functions: controllable locations p = 3 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  28. Operations on cost functions: controllable locations p = 3 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  29. Operations on cost functions: controllable locations p = 3 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  30. Operations on cost functions: controllable locations p = 3 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  31. Operations on cost functions: uncontrollable locations p = 2 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  32. Operations on cost functions: uncontrollable locations p = 2 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  33. Operations on cost functions: uncontrollable locations p = 2 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  34. Operations on cost functions: uncontrollable locations p = 2 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  35. Operations on cost functions: uncontrollable locations p = 2 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  36. Operations on cost functions: uncontrollable locations p = 2 ˙ p = 5 ˙ p = 3 ˙ p = 2 ˙ p = 1 ˙

  37. Inductive proof Ideas of the proof Induction on the number of non-urgent locations in the SCC base cases: all locations are urgent (thus uncontrollable); there is only one location, which is controllable (thus non-urgent). induction step: we consider one of the non-urgent locations having minimal cost rate: if it is controllable, we create two SCCs having one less non-urgent location; if it is uncontrollable, we make it urgent and add an extra outside cost function to which it can go. Skip proof

  38. Inductive proof – base cases p =1 ˙ p =3 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  39. Inductive proof – base cases p =1 ˙ p =3 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  40. Inductive proof – base cases p =1 ˙ p =3 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  41. Inductive proof – base cases p =3 ˙

  42. Inductive proof – base cases p =3 ˙

  43. Inductive proof – base cases p =3 ˙

  44. Inductive proof – base cases p =3 ˙

  45. Inductive proof – base cases p =3 ˙

  46. Inductive proof – inductive cases When q min is controllable: p =3 ˙ p =5 ˙ x ≤ 1 p =1 ˙ p =2 ˙

  47. Inductive proof – inductive cases When q min is controllable: Let σ be a winning strategy. p =3 ˙ p =5 ˙ x ≤ 1 p =1 ˙ p =2 ˙

  48. Inductive proof – inductive cases When q min is controllable: Let σ be a winning strategy. Assume there exists an outcome of σ s.t.: p =3 ˙ p =5 ˙ ( q min , u ) → ∗ ( q min , v ) → ∗ win x ≤ 1 with 0 ≤ u < v ≤ 1. p =1 ˙ p =2 ˙

  49. Inductive proof – inductive cases When q min is controllable: Let σ be a winning strategy. Assume there exists an outcome of σ s.t.: p =3 ˙ p =5 ˙ ( q min , u ) → ∗ ( q min , v ) → ∗ win x ≤ 1 with 0 ≤ u < v ≤ 1. p =1 ˙ p =2 ˙ Then σ is not optimal: waiting in q min would have been cheaper.

  50. Inductive proof – inductive cases When q min is controllable: p =3 ˙ p =5 ˙ x ≤ 1 p =1 ˙ p =2 ˙

  51. Inductive proof – inductive cases When q min is controllable: p =3 ˙ p =5 ˙ p =3 ˙ p =5 ˙ x ≤ 1 x ≤ 1 p =1 ˙ p =1 ˙ p =2 ˙ p =2 ˙

  52. Inductive proof – inductive cases When q min is controllable: p =3 ˙ p =5 ˙ p =3 ˙ p =5 ˙ x ≤ 1 x ≤ 1 + ∞ p =2 ˙ p =2 ˙ ˙ p =1

  53. Inductive proof – inductive cases When q min is controllable: m n q min G m 1 n 1 G ′ q min m 2 + ∞ n 2

  54. Inductive proof – inductive cases When q min is controllable: m Theorem OptCost G ′ ( q 1 , x ) = OptCost G ( q , x ) . n q min G m 1 n 1 G ′ q min m 2 + ∞ n 2

  55. Inductive proof – inductive cases When q min is controllable: m Theorem OptCost G ′ ( q 1 , x ) = OptCost G ( q , x ) . n q min G Theorem Let σ ′ be an ( ε ′ , N ′ )-acceptable strategy m 1 for G ′ . Let  σ ′ ( q 2 , x )  n 1  G ′ σ ( q , x ) = if Cost G ′ ( q 2 , x ) ≤ OptCost G ′ ( q min , x )  σ ′ ( q 1 , x ) otherwise  q min m 2 Then σ is (3 ε ′ , N )-acceptable in G , for + ∞ n 2 some N independant of ε ′ .

  56. Inductive proof – inductive cases When q min is uncontrollable: p =3 ˙ p =1 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  57. Inductive proof – inductive cases When q min is uncontrollable: Make q min urgent and apply I.H.: p =3 ˙ p =1 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  58. Inductive proof – inductive cases When q min is uncontrollable: Make q min urgent and apply I.H.: p =3 ˙ p =1 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  59. Inductive proof – inductive cases When q min is uncontrollable: Make q min urgent and apply I.H.: p =3 ˙ p =1 ˙ First instance where slope x ≤ 1 less than c min p =5 ˙ p =1 ˙

  60. Inductive proof – inductive cases When q min is uncontrollable: Make q min urgent and apply I.H.: p =3 ˙ p =1 ˙ It’s better to x ≤ 1 wait in q min ... p =5 ˙ p =1 ˙

  61. Inductive proof – inductive cases When q min is uncontrollable: Make q min urgent and apply I.H.: p =3 ˙ p =1 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  62. Inductive proof – inductive cases When q min is uncontrollable: Apply I.H. again: p =3 ˙ p =1 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  63. Inductive proof – inductive cases When q min is uncontrollable: Apply I.H. again: p =3 ˙ p =1 ˙ x ≤ 1 p =5 ˙ p =1 ˙

  64. Inductive proof – inductive cases When q min is uncontrollable: Apply I.H. again: p =3 ˙ p =1 ˙ First instance where slope x ≤ 1 less than c min p =5 ˙ p =1 ˙

  65. Inductive proof – inductive cases When q min is uncontrollable: Apply I.H. again: p =3 ˙ p =1 ˙ It’s better to wait in q min ... x ≤ 1 p =5 ˙ p =1 ˙

  66. Inductive proof – inductive cases When q min is uncontrollable: p =3 ˙ p =1 ˙ This procedure terminates because x ≤ 1 fragments having slope strictly less than c min are fragments of outside p =5 ˙ p =1 ˙ functions.

  67. Outline of the talk Introduction 1 Definitions and examples 2 Existence of optimal strategies in 1PTGAs is decidable 3 (Pseudo-)algorithm for computing the optimal cost 4 Conclusion 5

Recommend


More recommend