control of uncertain hybrid nonlinear systems using particle filters
DESCRIPTION
This paper proposes an optimization-based algorithm for the control of uncertain hybrid nonlinear systems. The considered system class combines the nondeterministic evolution of a discrete-time Markov process with the deterministic switching of continuous dynamics which itself contains uncertain elements. A weighted particle filter approach is used to approximate the uncertain evolution of the system by a set of deterministic runs. The desired control performance for a finite time horizon is encoded by a suitable cost function and a chance-constraint, which restricts the maximum probability for entering unsafe state sets. The optimization considers input and state constraints in addition. It is demonstrated that the resulting optimization problem can be solved by techniques of conventional mixed-integer nonlinear programming (MINLP). As an illustrative example, a path planning scenario of a ground vehicle with switching nonlinear dynamics is presented.TRANSCRIPT
Control of Uncertain Hybrid Nonlinear SystemsUsing Particle Filters
Leonhard Asselborn Martin Jilg Olaf Stursberg
Institute of Control and System TheoryUniversity of Kassel
www.control.eecs.uni-kassel.de
Introduction
• Considered class of models: hybrid nonlinear system with deterministic andnondeterministic transitions and uncertain continuous dynamics.
• Formulation of a point-to-region open-loop optimal control problem.• Contribution: proposal of a particle-filter based solution and feasibility
study for an example.
Introduction Model Definition Problem Setup Method Example Conclusion 2
Introduction
• Considered class of models: hybrid nonlinear system with deterministic andnondeterministic transitions and uncertain continuous dynamics.
• Formulation of a point-to-region open-loop optimal control problem.• Contribution: proposal of a particle-filter based solution and feasibility
study for an example.
x1
x2
Region 1 Region 2
t0
t1 t2te
t3
t4
Unsafe Region
Goal Region
initial state
uncertain hybrid trajectory
Introduction Model Definition Problem Setup Method Example Conclusion 2
Stochastic Hybrid Models (1)
• Stochastic hybrid models are a suitable tool for a wide range of processes
Introduction Model Definition Problem Setup Method Example Conclusion 3
Stochastic Hybrid Models (1)
• Stochastic hybrid models are a suitable tool for a wide range of processes
• Three main representations of Stochastic hybrid models in literature1. Stochastic Hybrid Systems (SHS)
◮ Randomness only in the continuous dynamics
2. Switching Diffusion Processes (SDP)◮ Random cont. dynamics and spontaneous transitions according to a Poisson
process
3. Piecewise Deterministic Markov Processes (PDMP)◮ Deterministic continuous dynamics and spontaneous or autonomous transitions
according to a Poisson Process and state space partitions, respectively.
Introduction Model Definition Problem Setup Method Example Conclusion 3
Stochastic Hybrid Models (2)
Control methods for Stochastic Hybrid Models in the literature (as far asrelevant for this paper):
Method specification Model specification
optimalcontrol
particlefilter
chanceconstraints
uncertaindynamics
nonlineardynamics
stochasticevents
deterministicevents
[1] X - X - - X X
[2] X X - X X X -
[3] X - - - X X -
[4] X - X X - - -
[5] - X - X X X -
[6] X - X X - - -
[1]: Bemporad et. al.: Model-Predictive Control of Discrete Hybrid Stochastic Automata, 2011,[2]: Blackmore et. al.: Optimal Robust Predictive Control of Nonlinear Systems under Probabilistic Uncertainty using Particles, 2007,[3]: Adamek et. al.: Stochastic Optimal Control for Hybrid Systems with Uncertain Dynamics, 2008,[4]: Ding et. al.: Increasing Efficiency of Optimization-based Path Planning for Robotic Manipulators, 2011,[5]: Li et. al.: Risk-Sensitive Cubature Filtering for Jump Markov Nonlinear Systems and Its Application to Land Vehicle Positioning, 2011,[6]: Vitus et. al.: Closed-Loop Belief Space Planning for Linear, Gaussian Systems, 2011
Introduction Model Definition Problem Setup Method Example Conclusion 4
Considered Class of Model
The “Jump Markov Nonlinear Hybrid Automaton (JMNHA)”:• No resets in the continuous state variable for autonomous transitions
based on a state space partition
• Spontaneous transitions according to a Markov Process.
• Uncertain nonlinear dynamics: deterministic behavior for t ∈]tk, tk+1[,stochastic pertubation at tk.
→ Model simplification by evaluating the nondeterministic events in discretetime.
Introduction Model Definition Problem Setup Method Example Conclusion 5
Considered Class of Model
The “Jump Markov Nonlinear Hybrid Automaton (JMNHA)”:• No resets in the continuous state variable for autonomous transitions
based on a state space partition
• Spontaneous transitions according to a Markov Process.
• Uncertain nonlinear dynamics: deterministic behavior for t ∈]tk, tk+1[,stochastic pertubation at tk.
PDMP
→ Model simplification by evaluating the nondeterministic events in discretetime.
Introduction Model Definition Problem Setup Method Example Conclusion 5
Considered Class of Model
The “Jump Markov Nonlinear Hybrid Automaton (JMNHA)”:• No resets in the continuous state variable for autonomous transitions
based on a state space partition
• Spontaneous transitions according to a Markov Process.
• Uncertain nonlinear dynamics: deterministic behavior for t ∈]tk, tk+1[,stochastic pertubation at tk.
PDMP
→ Model simplification by evaluating the nondeterministic events in discretetime.
Introduction Model Definition Problem Setup Method Example Conclusion 5
Considered Class of Model
The “Jump Markov Nonlinear Hybrid Automaton (JMNHA)”:• No resets in the continuous state variable for autonomous transitions
based on a state space partition
• Spontaneous transitions according to a Markov Process.
• Uncertain nonlinear dynamics: deterministic behavior for t ∈]tk, tk+1[,stochastic pertubation at tk.
PDMP
→ Model simplification by evaluating the nondeterministic events in discretetime.
Introduction Model Definition Problem Setup Method Example Conclusion 5
Considered Class of Model
The “Jump Markov Nonlinear Hybrid Automaton (JMNHA)”:• No resets in the continuous state variable for autonomous transitions
based on a state space partition
• Spontaneous transitions according to a Markov Process.
• Uncertain nonlinear dynamics: deterministic behavior for t ∈]tk, tk+1[,stochastic pertubation at tk.
PDMP
JMNHA
→ Model simplification by evaluating the nondeterministic events in discretetime.
Introduction Model Definition Problem Setup Method Example Conclusion 5
Considered Class of Model
The “Jump Markov Nonlinear Hybrid Automaton (JMNHA)”:• No resets in the continuous state variable for autonomous transitions
based on a state space partition
• Spontaneous transitions according to a Markov Process.
• Uncertain nonlinear dynamics: deterministic behavior for t ∈]tk, tk+1[,stochastic pertubation at tk.
PDMP
JMNHA
xk
xk+1
xk+1 = xk +∫ tk+1
tkf(τ)dτ + νk
→ Model simplification by evaluating the nondeterministic events in discretetime.
Introduction Model Definition Problem Setup Method Example Conclusion 5
A class of Stochastic Hybrid Systems
Definition 1:
A Jump Markov Nonlinear Hybrid Automaton is defined by
H = (T, Tk, Z,X,U,U , f, ψq, ψd, ν)
with t ∈ T , tk ∈ Tk
• nonlinear continuous dynamics x = f(x, u, d, q), x(t) ∈ X, u(t) ∈ U ,d(tk) ∈ D, q(tk) ∈ Q
• hybrid state space S = X × Z, x ∈ X, z = (d, q) ∈ Z
• update function ψq : X ×X → 2Q for the state space region
• update function ψd : D ×D → [0, 1]nd for the Markov process
• uncertainty ν in the continuous state variable x, xk+1 = xk+1 + νk+1
Introduction Model Definition Problem Setup Method Example Conclusion 6
Admissible behavior of the JMNHA for an example
Continuous dynamics: � f1 � f2 ⊲ f3 ⊲ f4
x1
x2
X1 X2
1
0.9
0.1 0.3
0.7
1
2
Markov process
• Two state space regions X1, X2 → Q = {1, 2}
• Markov Process with state set D = {1, 2}
• A sequence of hybrid states sk := s(tk) = ((dk, qk), xk) is denoted byφs = (s0, s1, s2, ...).
• A set of feasible runs of H for input functions U is denoted by Φs,U
Introduction Model Definition Problem Setup Method Example Conclusion 7
Admissible behavior of the JMNHA for an example
Continuous dynamics: � f1 � f2 ⊲ f3 ⊲ f4
x1
x2
X1 X2
s0
s1
s2tes3
s4
1
0.9
0.1 0.3
0.7
1
2
Markov process
• Two state space regions X1, X2 → Q = {1, 2}
• Markov Process with state set D = {1, 2}
• A sequence of hybrid states sk := s(tk) = ((dk, qk), xk) is denoted byφs = (s0, s1, s2, ...).
• A set of feasible runs of H for input functions U is denoted by Φs,U
Introduction Model Definition Problem Setup Method Example Conclusion 7
Robust Optimal Control Problem
The goal of the proposed method is to control H in an optimal manner w.r.t.chance, input and state constraints.
Introduction Model Definition Problem Setup Method Example Conclusion 8
Robust Optimal Control Problem
The goal of the proposed method is to control H in an optimal manner w.r.t.chance, input and state constraints.
• Perfomance index: Jφs=
∑nt
k=1 h(tk, xk, dk, qk, uk)
Introduction Model Definition Problem Setup Method Example Conclusion 8
Robust Optimal Control Problem
The goal of the proposed method is to control H in an optimal manner w.r.t.chance, input and state constraints.
• Perfomance index: Jφs=
∑nt
k=1 h(tk, xk, dk, qk, uk)
• Unsafe sets: A := ∪na
i=1Ai ⊂ X
Introduction Model Definition Problem Setup Method Example Conclusion 8
Robust Optimal Control Problem
The goal of the proposed method is to control H in an optimal manner w.r.t.chance, input and state constraints.
• Perfomance index: Jφs=
∑nt
k=1 h(tk, xk, dk, qk, uk)
• Unsafe sets: A := ∪na
i=1Ai ⊂ X
• A maximally permitted probability for entering an unsafe set: δ
Introduction Model Definition Problem Setup Method Example Conclusion 8
Robust Optimal Control Problem
The goal of the proposed method is to control H in an optimal manner w.r.t.chance, input and state constraints.
• Perfomance index: Jφs=
∑nt
k=1 h(tk, xk, dk, qk, uk)
• Unsafe sets: A := ∪na
i=1Ai ⊂ X
• A maximally permitted probability for entering an unsafe set: δ
• Goal set: G ⊂ X
Introduction Model Definition Problem Setup Method Example Conclusion 8
Robust Optimal Control Problem
The goal of the proposed method is to control H in an optimal manner w.r.t.chance, input and state constraints.
• Perfomance index: Jφs=
∑nt
k=1 h(tk, xk, dk, qk, uk)
• Unsafe sets: A := ∪na
i=1Ai ⊂ X
• A maximally permitted probability for entering an unsafe set: δ
• Goal set: G ⊂ X
Problem Definition
minuT∈U
E(Jφs)
s.t. s(t0) = ((d0, q0), x0 + ν0)
φs ∈ Φs,U
x(tf ) ∈ G
Prob(xT,φs∈ A for any t ∈ T ) ≤ δ
Introduction Model Definition Problem Setup Method Example Conclusion 8
Approximation by Particle Filters
The main properties of a Particle Filter 1
1[2]: L. Blackmore, B. C. Williams: Optimal Robust Predictive Control of Nonlinear Systemsunder Probabilistic Uncertainty using Particles, 2007
Introduction Model Definition Problem Setup Method Example Conclusion 9
Approximation by Particle Filters
The main properties of a Particle Filter 1
• A particle represents a sample of a random variable Ξ drawn from a givenprobability distribution:
{
ξ(1), . . . , ξ(N)}
• The expected value can be approximated by the sample mean:
E(Ξ) = 1N
∑Ni=1 ξ
(i)
• Each particle represents a deterministic realization of the JMNHA over a finitehorizon.
1[2]: L. Blackmore, B. C. Williams: Optimal Robust Predictive Control of Nonlinear Systemsunder Probabilistic Uncertainty using Particles, 2007
Introduction Model Definition Problem Setup Method Example Conclusion 9
Approximation by Particle Filters
The main properties of a Particle Filter 1
• A particle represents a sample of a random variable Ξ drawn from a givenprobability distribution:
{
ξ(1), . . . , ξ(N)}
• The expected value can be approximated by the sample mean:
E(Ξ) = 1N
∑Ni=1 ξ
(i)
• Each particle represents a deterministic realization of the JMNHA over a finitehorizon.
The particles are used to transform a stochastic optimization problem intoa deterministic variant.
1[2]: L. Blackmore, B. C. Williams: Optimal Robust Predictive Control of Nonlinear Systemsunder Probabilistic Uncertainty using Particles, 2007
Introduction Model Definition Problem Setup Method Example Conclusion 9
Approximation by Particle Filters
The main properties of a Particle Filter 1
• A particle represents a sample of a random variable Ξ drawn from a givenprobability distribution:
{
ξ(1), . . . , ξ(N)}
• The expected value can be approximated by the sample mean:
E(Ξ) = 1N
∑Ni=1 ξ
(i)
• Each particle represents a deterministic realization of the JMNHA over a finitehorizon.
The particles are used to transform a stochastic optimization problem intoa deterministic variant.
The setup of a chance constraint
• Indicator function 1A(x(i)1:nt
) denotes if the i-th trajectory is in A at any time.
• Binary vector λ ∈ {0, 1}N×1 is used for a mixed integer formulation:
PA = 1N
· w · λ ≤ δ with weighting vector w.
1[2]: L. Blackmore, B. C. Williams: Optimal Robust Predictive Control of Nonlinear Systemsunder Probabilistic Uncertainty using Particles, 2007
Introduction Model Definition Problem Setup Method Example Conclusion 9
Proposed Solution Scheme
x1
x2
X1 X2
s0
A1
G
Algorithm (1)
• Define and initialize H.
• Set up a suitable cost function h(tk, xk, dk, qk, uk), unsafe set A, goal setG, number of particles N and max. permitted probability δ.
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0
A1
G
Algorithm (2)
• Draw N samples d(i)0 , x
(i)0 according to to the corresponding probability
distribution.
• Generate a modified transition probability matrix according to theweighting concept and draw a sequence d
(i)1:nt
for each d(i)0 .
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0
A1
G
Algorithm (3)
• Generate a sequence of disturbance samples ν(i)1:nt
∼ N (µν , σν) andcalculate the weights wi.
• Choose an initial sequence of control inputs u0:nt−1 and compute the
future trajectories x(i)1:nt
for every particle.
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0 s1
A1
G
Algorithm (3)
• Generate a sequence of disturbance samples ν(i)1:nt
∼ N (µν , σν) andcalculate the weights wi.
• Choose an initial sequence of control inputs u0:nt−1 and compute the
future trajectories x(i)1:nt
for every particle.
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0 s1 s2te s3
A1
G
Algorithm (3)
• Generate a sequence of disturbance samples ν(i)1:nt
∼ N (µν , σν) andcalculate the weights wi.
• Choose an initial sequence of control inputs u0:nt−1 and compute the
future trajectories x(i)1:nt
for every particle.
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0 s1 s2te s3
A1
G
Algorithm (4)
• Formulate the chance constraint in terms of the weighted particles:PA = 1
N· w · λ ≤ δ
• Determine the approximated costs:
J =nt∑
k=1
Jk =nt∑
k=1
(
1N
N∑
i=1
wi · h(tk, x(i)k , uk, dk, qk)
)
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0 s1s2
te s3
A1
G
Algorithm (4)
• Formulate the chance constraint in terms of the weighted particles:PA = 1
N· w · λ ≤ δ
• Determine the approximated costs:
J =nt∑
k=1
Jk =nt∑
k=1
(
1N
N∑
i=1
wi · h(tk, x(i)k , uk, dk, qk)
)
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0 s1s2
te s3
A1
G
MINLP optimization problemu0:nt−1 = arg min
uT∈UJ
s.t. s(i)(t0) = ((d
(i)0 , q
(i)0 ), x
(i)0 + ν
(i)0 )
φ(i)s ∈ Φs,U , E(xnt
) ∈ G, PA =1
Nw · λ ≤ δ
Introduction Model Definition Problem Setup Method Example Conclusion 10
Proposed Solution Scheme
x1
x2
X1 X2
s0 s1s2
te s3
A1
G
MINLP optimization problemu0:nt−1 = arg min
uT∈UJ
s.t. s(i)(t0) = ((d
(i)0 , q
(i)0 ), x
(i)0 + ν
(i)0 )
φ(i)s ∈ Φs,U , E(xnt
) ∈ G, PA =1
Nw · λ ≤ δ
Introduction Model Definition Problem Setup Method Example Conclusion 10
Trajectory planning for a ground vehicle
• vehicle accelerates tomaximum speed
• vehicle turns left and staysclose to the obstacle
• vehicle decelerates in frontof the goal
• particles lead to manytrajectories with steeringand braking failure due tothe weighting concept
• at most 1 of the 10particle trajectories crossthe obstacle
• MINLP was solved with anSQP method within aBranch-and-Boundenvironment.
isocost lines
particles
goal
stop
start
trajectory
η
ζ
−2 0 2 4 6 8 10 12 14 16 18
5
10
15
20
25
Introduction Model Definition Problem Setup Method Example Conclusion 11
Conclusion
Conclusion
Summary:
• Introduction of an uncertain nonlinear hybrid system.
• Presentation of an algorithm for optimal open-loop control.
• A numerical example showed the results of the proposed method.
Remarks:
• Scheme generates a probabilistically safe sub-optimal trajectory
• Computationally expensive due to the complex model class and thechallenging problem.
• Branch-and-Bound may cut off branches which contain feasible solutions.
Future work:
• Efficient encoding of the chance constraints and alternatives for solvingthe optimization.
• Theoretical analysis concerning the convergence of the approximatedoptimization problem.
Introduction Model Definition Problem Setup Method Example Conclusion 12