Modeling and Analysis of Hybrid Systems

Size: px

Start display at page:

Download "Modeling and Analysis of Hybrid Systems"

Deirdre Hutchinson
5 years ago
Views:

1 Modeling and Analysis of Hybrid Systems Linear hybrid automata II: Approximation of reachable state sets Prof. Dr. Erika Ábrahám Informatik 2 - Theory of Hybrid Systems RWTH Aachen University SS 2015 Ábrahám - Hybrid Systems 1 / 25

2 We had a look at state set approximations by convex polyhedra, and at the basic operations testing for membership, intersection, and union on these. Thus we can approximate state sets and compute with them. How is all this used in the reachability analysis procedure? Ábrahám - Hybrid Systems 2 / 25

3 General reachability procedure Input: Set Init of initial states. Output: Set R of reachable states. Algorithm: R new := Init; R := ; while (R new ){ R := R R new ; R new := Reach (R new )\R; } What is Reach? Ábrahám - Hybrid Systems 3 / 25

4 What is Reach? For hybrid systems, independently of the exact definition of Reach, it will involve the following computations: Given a state set R, compute the set of states reachable from R by a flow (i.e., time transisiton), and the set of states reachable from R by a jump (i.e., discrete transition). Computing the jump successors of a set can be done with the operations we already introduced. The harder part is computing the flow successors. So let s have a look at that... Ábrahám - Hybrid Systems 4 / 25

5 Approximating a flow pipe Consider a dynamical system with state equation ẋ = f(x(t)). Ábrahám - Hybrid Systems 5 / 25

6 Approximating a flow pipe Consider a dynamical system with state equation ẋ = f(x(t)). We assume f to be Lipschitz continuous. Ábrahám - Hybrid Systems 5 / 25

7 Approximating a flow pipe Consider a dynamical system with state equation ẋ = f(x(t)). We assume f to be Lipschitz continuous. Wikipedia: Intuitively, a Lipschitz continuous function is limited in how fast it can change: for every pair of points on the graph of this function, the absolute value of the slope of the line connecting them is no greater than a definite real number [... ].. Ábrahám - Hybrid Systems 5 / 25

8 Approximating a flow pipe Consider a dynamical system with state equation ẋ = f(x(t)). We assume f to be Lipschitz continuous. Wikipedia: Intuitively, a Lipschitz continuous function is limited in how fast it can change: for every pair of points on the graph of this function, the absolute value of the slope of the line connecting them is no greater than a definite real number [... ].. Lipschitz continuity implies the existence and uniqueness of the solution to an initial value problem, i.e., for every initial state x 0 there is a unique solution x(t, x 0 ) to the state equation. Ábrahám - Hybrid Systems 5 / 25

9 Approximating a flow pipe The set of reachable states at time t from a set of initial states X 0 is defined as R t (X 0 ) = {x t x 0 X 0. x t = x(t, x 0 )}. Ábrahám - Hybrid Systems 6 / 25

10 Approximating a flow pipe The set of reachable states at time t from a set of initial states X 0 is defined as R t (X 0 ) = {x t x 0 X 0. x t = x(t, x 0 )}. The set of reachable states, the flow pipe, from X 0 in the time interval [0, t f ] is defined as R [0,tf ](X 0 ) = t [0,tf ]R t (X 0 ). Ábrahám - Hybrid Systems 6 / 25

11 Approximating a flow pipe The set of reachable states at time t from a set of initial states X 0 is defined as R t (X 0 ) = {x t x 0 X 0. x t = x(t, x 0 )}. The set of reachable states, the flow pipe, from X 0 in the time interval [0, t f ] is defined as R [0,tf ](X 0 ) = t [0,tf ]R t (X 0 ). We describe a solution which approximates the flow pipe by a sequence of convex polytopes. Ábrahám - Hybrid Systems 6 / 25

12 Problem statement for polyhedral approximation of flow pipes Given a set X 0 of initial states which is a polytope, and a final time t f, compute a polyhedral approximation ˆR [0,tf ](X 0 ) to the flow pipe R [0,tf ](X 0 ) such that R [0,tf ](X 0 ) ˆR [0,tf ](X 0 ). Ábrahám - Hybrid Systems 7 / 25

13 Flow pipe segmentation Since a single convex polyhedron would strongly overapproximate the flow pipe, we compute a sequence of convex polyhedra, each approximating a flow pipe segment. Ábrahám - Hybrid Systems 8 / 25

14 Segmented flow pipe approximation Let the time interval [0, t f ] be divided into 0 < N N time segments with t i = i tf N. [0, t 1 ], [t 1, t 2 ],..., [t N 1, t f ] Ábrahám - Hybrid Systems 9 / 25

15 Segmented flow pipe approximation Let the time interval [0, t f ] be divided into 0 < N N time segments with t i = i tf N. [0, t 1 ], [t 1, t 2 ],..., [t N 1, t f ] We generate an approximation ˆR [t1,t 2 ](X 0 ) for each flow pipe segment: R [t1,t 2 ](X 0 ) ˆR [t1,t 2 ](X 0 ). Ábrahám - Hybrid Systems 9 / 25

16 Segmented flow pipe approximation Let the time interval [0, t f ] be divided into 0 < N N time segments with t i = i tf N. [0, t 1 ], [t 1, t 2 ],..., [t N 1, t f ] We generate an approximation ˆR [t1,t 2 ](X 0 ) for each flow pipe segment: R [t1,t 2 ](X 0 ) ˆR [t1,t 2 ](X 0 ). The complete flow pipe approximation is the union of the approximation of all N pipe segments: R [0,tf ](X 0 ) ˆR [0,tf ](X 0 ) = ˆR [tk 1,t k ](X 0 ) k=1,...,n Ábrahám - Hybrid Systems 9 / 25

17 Approaches Next we discuss two possible approaches for flow pipe approximation, but there are different other techniques, too. Ábrahám - Hybrid Systems 10 / 25

18 The first approach Ábrahám - Hybrid Systems 11 / 25

19 Linear hybrid automata II: Time evolution Ábrahám - Hybrid Systems 12 / 25

20 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Ábrahám - Hybrid Systems 12 / 25

21 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu R [0,δ] R [δ,2δ] R [2δ,3δ] Ábrahám - Hybrid Systems 12 / 25

22 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i Ω 0 R [0,δ] R [δ,2δ] Ω 1 Ω 2 R [2δ,3δ] Ábrahám - Hybrid Systems 12 / 25

23 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: Ábrahám - Hybrid Systems 12 / 25

24 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: Ábrahám - Hybrid Systems 12 / 25

25 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

26 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

27 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

28 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

29 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

30 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

31 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

32 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

33 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

34 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: Ω 0 e Aδ V 0 0 δ 2δ t V 0 Ábrahám - Hybrid Systems 12 / 25

35 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: Ω 0 0 δ 2δ t Ábrahám - Hybrid Systems 12 / 25

36 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: The remaining ones: Ω 0 0 δ 2δ t Ábrahám - Hybrid Systems 12 / 25

37 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: The remaining ones: Ω 0 0 δ 2δ t e Aδ Ω 0 Ábrahám - Hybrid Systems 12 / 25

38 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: The remaining ones: Ω 0 0 δ 2δ t e Aδ Ω 0 Ábrahám - Hybrid Systems 12 / 25

39 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: The remaining ones: Ω 0 0 δ 2δ t e Aδ Ω 0 Ábrahám - Hybrid Systems 12 / 25

40 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: The remaining ones: Ω 0 0 δ 2δ t Ábrahám - Hybrid Systems 12 / 25

41 Linear hybrid automata II: Time evolution Assume ẋ = Ax + Bu Compute Ω 0, Ω 1,... such that R [iδ,(i+1)δ] Ω i The first flowpipe segment: The remaining ones: Ω 0 e Aδ Ω 0 V 0 δ 2δ t Ábrahám - Hybrid Systems 12 / 25

42 Linear hybrid automata II: Discrete steps (jumps) Ω 3 Ω 2 Ω 1 Ω 0 Ábrahám - Hybrid Systems 13 / 25

43 Linear hybrid automata II: Discrete steps (jumps) Ω 3 Ω 2 Ω 1 Ω 0 Ábrahám - Hybrid Systems 13 / 25

44 Linear hybrid automata II: Discrete steps (jumps) Ω 3 Ω 2 Ω 1 Ω 0 Ábrahám - Hybrid Systems 13 / 25

45 Linear hybrid automata II: Discrete steps (jumps) Ω 3 Ω 1 Ω 2 Ábrahám - Hybrid Systems 13 / 25

46 Linear hybrid automata II: Discrete steps (jumps) Ω 3 Π 3 Ω 1 Ω 2 Π 1 Π 2 Ábrahám - Hybrid Systems 13 / 25

47 Linear hybrid automata II: Discrete steps (jumps) V 1 Ω 3 Π 3 Ω 1 Ω 2 Π 1 Π 2 Ábrahám - Hybrid Systems 13 / 25

48 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

49 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

50 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

51 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

52 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

53 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

54 Linear hybrid automata II: The global picture Ábrahám - Hybrid Systems 14 / 25

55 The second approach Ábrahám - Hybrid Systems 15 / 25

56 Literatur Alongkrit Chutinan and Bruce H. Krogh: Computing Polyhedral Approximations to Flow Pipes for Dynamic Systems In Proceedings of the 37rd IEEE Conference on Decision and Control, 1998 Olaf Stursberg and Bruce H. Krogh: Efficient Representation and Computation of Reachable Sets for Hybrid Systems Hybrid Systems: Computation and Control, LNCS 2623, pp , 2003 Ábrahám - Hybrid Systems 16 / 25

57 Some notations We will use the following notations: Let POLY (C, d) denote the convex polytope defined by the pair (C, d) R m n R m according to POLY (C, d) = {x Cx d}. For a polytope P by V (P ) we denote the finite set of its vertices, which are points in P that cannot be written as a strict convex combination of any other two points in P. Given a finite set of points Γ, the convex hull conv(γ) of Γ is the smallest convex set that contains Γ. Ábrahám - Hybrid Systems 17 / 25

58 Approximation of a flow pipe segment The approximation of the flow pipe for the time segment [t k 1, t k ] (k {1,..., N}) consists of the following steps: Ábrahám - Hybrid Systems 18 / 25

59 Approximation of a flow pipe segment The approximation of the flow pipe for the time segment [t k 1, t k ] (k {1,..., N}) consists of the following steps: Evolve vertices: Compute the set of points reachable from the vertices of X 0 in time t i 1 and in time t i. Ábrahám - Hybrid Systems 18 / 25

60 Approximation of a flow pipe segment The approximation of the flow pipe for the time segment [t k 1, t k ] (k {1,..., N}) consists of the following steps: Evolve vertices: Compute the set of points reachable from the vertices of X 0 in time t i 1 and in time t i. Determine hull: Compute the convex hull of those points. Ábrahám - Hybrid Systems 18 / 25

61 Approximation of a flow pipe segment The approximation of the flow pipe for the time segment [t k 1, t k ] (k {1,..., N}) consists of the following steps: Evolve vertices: Compute the set of points reachable from the vertices of X 0 in time t i 1 and in time t i. Determine hull: Compute the convex hull of those points. Bloat hull: Enlarge the hull until it contains all points of the flow pipe segment. Ábrahám - Hybrid Systems 18 / 25

62 1. Evolve vertices To gain some geometrical information about the flow pipe segment, we begin with taking sample points at times t k 1 and t k from the trajectories emanating from the vertices of X 0. Ábrahám - Hybrid Systems 19 / 25

63 1. Evolve vertices To gain some geometrical information about the flow pipe segment, we begin with taking sample points at times t k 1 and t k from the trajectories emanating from the vertices of X 0. In particular, we compute the sets V tk 1 (X 0 ) and V tk (X 0 ) where V t (X 0 ) = {x(t, v) v V (X 0 )}. Ábrahám - Hybrid Systems 19 / 25

64 1. Evolve vertices To gain some geometrical information about the flow pipe segment, we begin with taking sample points at times t k 1 and t k from the trajectories emanating from the vertices of X 0. In particular, we compute the sets V tk 1 (X 0 ) and V tk (X 0 ) where V t (X 0 ) = {x(t, v) v V (X 0 )}. Each point in the above sets can be obtained by analytic solution of the state equation and computing the value, or by simulation. Ábrahám - Hybrid Systems 19 / 25

65 2. Determine hull We use the evolved vertices in V tk 1 (X 0 ) and V tk (X 0 ) to form a convex hull which serves as an initial approximation to the flow pipe segment R [tk 1,t k ](X 0 ), denoted by Φ [tk 1,t k ](X 0 ) = conv(v tk 1 (X 0 ) V tk (X 0 )). Ábrahám - Hybrid Systems 20 / 25

66 2. Determine hull We use the evolved vertices in V tk 1 (X 0 ) and V tk (X 0 ) to form a convex hull which serves as an initial approximation to the flow pipe segment R [tk 1,t k ](X 0 ), denoted by Φ [tk 1,t k ](X 0 ) = conv(v tk 1 (X 0 ) V tk (X 0 )). Note that Φ [tk 1,t k ](X 0 ) may not contain the whole flow pipe segment R [tk 1,t k ](X 0 ). Ábrahám - Hybrid Systems 20 / 25

67 2. Determine hull We use the evolved vertices in V tk 1 (X 0 ) and V tk (X 0 ) to form a convex hull which serves as an initial approximation to the flow pipe segment R [tk 1,t k ](X 0 ), denoted by Φ [tk 1,t k ](X 0 ) = conv(v tk 1 (X 0 ) V tk (X 0 )). Note that Φ [tk 1,t k ](X 0 ) may not contain the whole flow pipe segment R [tk 1,t k ](X 0 ). Let (C Φ, d Φ ) be the matrix-vector pair defining the convex hull, i.e., Φ [tk 1,t k ](X 0 ) = POLY (C Φ, d Φ ). Ábrahám - Hybrid Systems 20 / 25

68 3. Bloat hull The normal vector on each face of the polytope points outward. Ábrahám - Hybrid Systems 21 / 25

69 3. Bloat hull The normal vector on each face of the polytope points outward. We use the normal vectors to the faces of this convex hull as a set of direction vectors to bloat the convex set until it contains the whole flow pipe segment. Ábrahám - Hybrid Systems 21 / 25

70 3. Bloat hull The normal vector on each face of the polytope points outward. We use the normal vectors to the faces of this convex hull as a set of direction vectors to bloat the convex set until it contains the whole flow pipe segment. Given: POLY (C Φ, d Φ ). Ábrahám - Hybrid Systems 21 / 25

71 3. Bloat hull The normal vector on each face of the polytope points outward. We use the normal vectors to the faces of this convex hull as a set of direction vectors to bloat the convex set until it contains the whole flow pipe segment. Given: POLY (C Φ, d Φ ). We want: R [tk 1,t k ](X 0 ) POLY (C Φ, d ). Ábrahám - Hybrid Systems 21 / 25

72 3. Bloat hull We compute d as the solution to the following optimization problem: min d volume[poly (C Φ, d)] (1) s.t. R [tk 1,t k ](X 0 ) POLY (C Φ, d). Ábrahám - Hybrid Systems 22 / 25

73 3. Bloat hull We compute d as the solution to the following optimization problem: min d volume[poly (C Φ, d)] (1) s.t. R [tk 1,t k ](X 0 ) POLY (C Φ, d). The ith component d i of the optimum d can be found by solving max x c T i x s.t. x R [tk 1,t k ](X 0 ). (2) Ábrahám - Hybrid Systems 22 / 25

74 3. Bloat hull We compute d as the solution to the following optimization problem: min d volume[poly (C Φ, d)] (1) s.t. R [tk 1,t k ](X 0 ) POLY (C Φ, d). The ith component d i of the optimum d can be found by solving max x c T i x s.t. x R [tk 1,t k ](X 0 ). (2) or, equivalently, max x 0,t c T i x(t, x 0 ) s.t. x 0 X 0, t [t k 1, t k ]. (3) Ábrahám - Hybrid Systems 22 / 25

75 3. Bloat hull We compute d as the solution to the following optimization problem: min d volume[poly (C Φ, d)] (1) s.t. R [tk 1,t k ](X 0 ) POLY (C Φ, d). The ith component d i of the optimum d can be found by solving max x c T i x s.t. x R [tk 1,t k ](X 0 ). (2) or, equivalently, max x 0,t c T i x(t, x 0 ) s.t. x 0 X 0, t [t k 1, t k ]. (3) Solution (x 0, t ) Solution x(t, x 0 ) to 3 to 2 Solution d i = ct i x(t, x 0 ) to 1. Ábrahám - Hybrid Systems 22 / 25

76 Example Van der Pol equation: ẋ 1 = x 2 ẋ 2 = 0.2(x 2 1 1)x 2 x 1. Intial set: X 0 = {(x 1, x 2 ) 0.8 x 1 1 x 2 = 0}. Time: t f = 10. Segments: 20 Ábrahám - Hybrid Systems 23 / 25

77 Other geometries for approximation Van der Pol equation with a third variable being a clock. Approximation with convex polyhedra and with oriented rectangular hull: Ábrahám - Hybrid Systems 24 / 25

78 Partitioning the initial set Var der Pol system with initial set X 0 = {(x 1, x 2 ) 5 x 1 45 x 2 = 0}. Ábrahám - Hybrid Systems 25 / 25

Modeling and Analysis of Hybrid Systems

Modeling and Analysis of Hybrid Systems 7. Linear hybrid automata II Prof. Dr. Erika Ábrahám Informatik 2 - LuFG Theory of Hybrid Systems RWTH Aachen University Szeged, Hungary, 27 September - 6 October