def iv_P_norm_expm(P_sqrt_T, M1, A, M2, tau):
"""
Bound on P-ellipsoid norm of (M1 (expm(A*t) - I) M2) for |t| < tau
using the theorem in arXiv:1911.02537, section "Norm bounding of summands"
@param P_sqrt_T: see iv_P_norm()
"""
P_sqrt_T = iv.matrix(P_sqrt_T)
M1 = iv.matrix(M1)
A = iv.matrix(A)
M2 = iv.matrix(M2)
# coerce tau to maximum
tau = abs(iv.mpf(tau)).b
# P-ellipsoid norms
M1_p = iv_P_norm(M=M1, P_sqrt_T=P_sqrt_T)
M2_p = iv_P_norm(M=M2, P_sqrt_T=P_sqrt_T)
A_p = iv_P_norm(M=A, P_sqrt_T=P_sqrt_T)
# A_pow[i] = A ** i
A_pow = _iv_matrix_powers(A)
# Work around bug in mpmath, see comment in iv_P_norm()
zero = iv.matrix(mp.zeros(len(A)))
M1 = zero + M1
# terms from [arXiv:1911.02537]
M1_Ai_M2_p = lambda i: iv_P_norm(M=M1 @ A_pow[i] @ M2, P_sqrt_T=P_sqrt_T)
gamma = lambda i: 1 / math.factorial(i) * (M1_Ai_M2_p(i) - M1_p * A_p ** i * M2_p)
max_norm = sum([gamma(i) * (tau ** i) for i in range(1, IV_NORM_EVAL_ORDER + 1)]) + M1_p * M2_p * (iv.exp(A_p * tau) - 1)
# the lower bound is always 0 (for t=0)
return mp.mpi([0, max_norm.b])
def test_terminal(self):
interval = iv.mpf([-0.005, 0.005])
print(interval < 0)
print(interval < 1)
print(interval > -1)
print(True or None)
print(False or None)
def reset(self):
self.state = tuple([iv.mpf([-0.005, 0.005]) for x in range(4)])
self.steps_beyond_done = None
return HyperRectangle.from_numpy(
np.array(
tuple([(float(x.a), float(x.b))
for i, x in enumerate(np.array(self.state))
])).transpose())
def iv_spectral_norm_rough(M):
"""
Fast but rough interval bound of spectral norm.
0 <= spectral_norm(M) <= sqrt(sum of all m[i,k]^2)
"""
norm = iv.norm(M, 2)
return iv.mpf([0, norm.b])
def assertInInterval(self, value, interval, relativeTolerance=0):
"""
assert that value is in the given interval
@param relativeTolerance: widen the interval by this factor
"""
self.assertIn(
value,
interval * (1 + iv.mpf([-1, +1]) * mp.mpf(relativeTolerance)))
def __init__(self,a,b,dim=False,index=None):
"""Class for estimating the maximum curvature.
Parameters
----------
a (int) - the starting value of the interval
b (int) - the ending value of the interval
dim (bool or int) - False if this is not an interval for a dimension
integer indicating the number of dimensions
index (int) - defines which dimension this interval corresponds to
"""
self.iv = iv.mpf([a,b])
self.iv_lambda = iv.mpf([0,0])
if dim:
assert isinstance(dim, int)
assert isinstance(index, int) and 0<=index<dim
self.iv_prime = np.array([iv.mpf([0,0]) for _ in range(dim)])
self.iv_prime[index] = iv.mpf([1,1])
else:
self.iv_prime = iv.mpf([0,0])
def from_bounds(cls, lower=None, upper=None, lower_upper=None):
"""
construct interval matrix from elementwise upper and lower bound
either given as separate vectors `lower` and `upper`, or as 2-by-n numpy array lower_upper.
"""
assert ((lower is None and upper is None) != (lower_upper is None))
if lower_upper is not None:
assert lower_upper.shape == (len(lower_upper), 2)
lower = lower_upper[:, 0]
upper = lower_upper[:, 1]
return cls.convert(lower) + (cls.convert(upper) -
cls.convert(lower)) * iv.mpf([0, 1])
def example_matrices(self, include_singular=True, random=100):
examples = [
iv.matrix([[1, 2], [4, 5]]) + iv.mpf([-1, +1]) * mp.mpf(1e-10),
iv.eye(2) * 0.5,
iv.diag([1, 1e-10]),
iv.eye(2) * 1e-302
]
for n in [1, 4, 17]:
for i in range(random // n):
examples.append(iv.matrix(mp.randmatrix(n)))
if include_singular:
examples.append(iv.matrix(mp.zeros(4)))
examples.append(iv.diag([1, 1e-200]))
return examples
def iv_spectral_norm(M):
"""
Good interval bound of spectral norm of a (interval) matrix
Theorem 3.2 from
Siegfried M. Rump. “Verified bounds for singular values, in particular for the
spectral norm of a matrix and its inverse”. In: BIT Numerical Mathematics 51.2
(Nov. 2010), pp. 367–384. DOI : 10.1007/s10543-010-0294-0.
"""
M = iv.matrix(M)
# imprecise SVD of M (no requirement on precision)
# _, _, V_T = mp.svd(iv_matrix_mid_to_mp(M)) # <-- configurable precision
_, _, V_T = scipy.linalg.svd(iv_matrix_mid_to_numpy_ndarray(M)) # <-- faster
# in [Rump 2010], SVD is defined as M = U @ S @ V.T,
# in mpmath it is M = U @ S @ V (not transposed) for U,S,V=svd(M)
# in scipy it is effectively the same as in mpmath, M = U @ S @ Vh for U,S,Vh = svd(M)
V = numpy_ndarray_to_mp_matrix(V_T).T
# now, everything is named as in [Rump2010], except that here, A is called M.
# all following computations are interval bounds
V = iv.matrix(V)
B = M @ V
BTB = B.T @ B
# split BTB such that DE = diagonal D + rest E
D = BTB @ 0
E = BTB @ 0
for i in range(BTB.rows):
for j in range(BTB.cols):
if i == j:
D[i,j] = BTB[i,j]
else:
E[i,j] = BTB[i,j]
# upper bound of spectral norm of I - V.T @ V
alpha = iv_spectral_norm_rough(iv.eye(len(M)) - V.T @ V)
# upper bound of spectral norm of E
epsilon = iv_spectral_norm_rough(E)
# maximum of D[i,i] (which are always >= 0)
d_max = iv.norm(D, mp.inf)
if alpha.b >= 1:
# this shouldn't happen - even an imprecise SVD will roughly have V.T @ V = I.
raise scipy.linalg.LinAlgError("Something's numerically wrong - the singular vectors are far from orthonormal")
# should this ever happen in reality, a valid return value would be:
# return iv.mpf([0, mp.inf])
try:
lower_bound = iv.sqrt((d_max - epsilon) / (1 + alpha)).a
except mp.libmp.libmpf.ComplexResult:
lower_bound=0;
# note that d_max, epsilon,alpha are intervals, so everything in the following computation is interval arithmetic
return iv.mpf([lower_bound, iv.sqrt((d_max + epsilon) / (1 - alpha)).b])
def apply(self, n, b, evaluation):
'IntegerLength[n_, b_]'
# Use interval arithmetic to account for "right" rounding
n, b = n.get_int_value(), b.get_int_value()
if n is None or b is None:
evaluation.message('IntegerLength', 'int')
return
if b <= 1:
evaluation.message('IntegerLength', 'base', b)
return
result = mp.mpf(iv.log(iv.mpf(mpz2mpmath(abs(n))), mpz2mpmath(b)).b)
result = mpz(mpmath2gmpy(result)) + 1
return Integer(result)
def approx_P_norm_expm(P_sqrt_T, M1, A, M2, tau):
"""
approximate max ( P_norm( M1 (expm(A*t) - I) M2 ) for |t| < tau )
@param P_sqrt_T: see iv_P_norm()
"""
P_sqrt_T = iv_matrix_mid_to_numpy_ndarray(P_sqrt_T)
M1 = iv_matrix_mid_to_numpy_ndarray(M1)
A = iv_matrix_mid_to_numpy_ndarray(A)
M2 = iv_matrix_mid_to_numpy_ndarray(M2)
# coerce tau to maximum
tau = float(mp.mpf(abs(iv.mpf(tau)).b))
max_norm = 0
for t in numpy.linspace(-tau, tau, 100):
matrix = numpy.matmul(M1, numpy.matmul(scipy.linalg.expm(A*t) - numpy.eye(len(A)), M2))
max_norm = max(max_norm, approx_P_norm(M=matrix, P_sqrt_T=P_sqrt_T))
return max_norm
def Interpolate1DFixedDuration(x0, x1, v0, v1, newDuration, vm, am):
x0 = ConvertFloatToMPF(x0)
x1 = ConvertFloatToMPF(x1)
v0 = ConvertFloatToMPF(v0)
v1 = ConvertFloatToMPF(v1)
vm = ConvertFloatToMPF(vm)
am = ConvertFloatToMPF(am)
newDuration = ConvertFloatToMPF(newDuration)
log.debug("\nx0 = {0}; x1 = {1}; v0 = {2}; v1 = {3}; vm = {4}; am = {5}; newDuration = {6}".\
format(mp.nstr(x0, n=_prec), mp.nstr(x1, n=_prec), mp.nstr(v0, n=_prec), mp.nstr(v1, n=_prec),
mp.nstr(vm, n=_prec), mp.nstr(am, n=_prec), mp.nstr(newDuration, n=_prec)))
# Check inputs
assert (vm > zero)
assert (am > zero)
if (newDuration < -epsilon):
return ParabolicCurve()
if (newDuration <= epsilon):
# Check if this is a stationary trajectory
if (FuzzyEquals(x0, x1, epsilon) and FuzzyEquals(v0, v1, epsilon)):
ramp0 = Ramp(v0, 0, 0, x0)
newCurve = ParabolicCurve(ramp0)
return newCurve
else:
# newDuration is too short to any movement to be made
return ParabolicCurve()
d = Sub(x1, x0)
# First assume no velocity bound -> re-interpolated trajectory will have only two ramps.
# Solve for a0 and a1 (the acceleration of the first and the last ramps).
# a0 = A + B/t0
# a1 = A + B/(t - t0)
# where t is the (new) total duration, t0 is the (new) duration of the first ramp, and
# A = (v1 - v0)/t
# B = (2d/t) - (v0 + v1).
newDurInverse = mp.fdiv(one, newDuration)
A = Mul(Sub(v1, v0), newDurInverse)
B = Sub(Prod([mp.mpf('2'), d, newDurInverse]), Add(v0, v1))
interval0 = iv.mpf([zero, newDuration]) # initial interval for t0
# Now consider the interval(s) computed from a0's constraints
sum1 = Neg(Add(am, A))
sum2 = Sub(am, A)
C = mp.fdiv(B, sum1)
D = mp.fdiv(B, sum2)
log.debug("\nA = {0}; \nB = {1}; \nC = {2}; \nD = {3}; \nsum1 = {4}; \nsum2 = {5};".\
format(mp.nstr(A, n=_prec), mp.nstr(B, n=_prec), mp.nstr(C, n=_prec), mp.nstr(D, n=_prec),
mp.nstr(sum1, n=_prec), mp.nstr(sum2, n=_prec)))
if (sum1 > zero):
# This implied that the duration is too short
log.debug("the given duration ({0}) is too short.".format(newDuration))
return ParabolicCurve()
if (sum2 < zero):
# This implied that the duration is too short
log.debug("the given duration ({0}) is too short.".format(newDuration))
return ParabolicCurve()
if IsEqual(sum1, zero):
raise NotImplementedError # not yet considered
elif sum1 > epsilon:
log.debug("sum1 > 0. This implies that newDuration is too short.")
return ParabolicCurve()
else:
interval1 = iv.mpf([C, inf])
if IsEqual(sum2, zero):
raise NotImplementedError # not yet considered
elif sum2 > epsilon:
interval2 = iv.mpf([D, inf])
else:
log.debug("sum2 < 0. This implies that newDuration is too short.")
return ParabolicCurve()
if Sub(interval2.a, interval1.b) > epsilon or Sub(interval1.a,
interval2.b) > epsilon:
# interval1 and interval2 do not intersect each other
return ParabolicCurve()
# interval3 = interval1 \cap interval2 : valid interval for t0 computed from a0's constraints
interval3 = iv.mpf(
[max(interval1.a, interval2.a),
min(interval1.b, interval2.b)])
# Now consider the interval(s) computed from a1's constraints
if IsEqual(sum1, zero):
raise NotImplementedError # not yet considered
elif sum1 > epsilon:
log.debug("sum1 > 0. This implies that newDuration is too short.")
return ParabolicCurve()
else:
interval4 = iv.mpf([Neg(inf), Add(C, newDuration)])
if IsEqual(sum2, zero):
raise NotImplementedError # not yet considered
elif sum2 > epsilon:
interval5 = iv.mpf([Neg(inf), Add(D, newDuration)])
else:
log.debug("sum2 < 0. This implies that newDuration is too short.")
return ParabolicCurve()
if Sub(interval5.a, interval4.b) > epsilon or Sub(interval4.a,
interval5.b) > epsilon:
log.debug("interval4 and interval5 do not intersect each other")
return ParabolicCurve()
# interval6 = interval4 \cap interval5 : valid interval for t0 computed from a1's constraints
interval6 = iv.mpf(
[max(interval4.a, interval5.a),
min(interval4.b, interval5.b)])
# import IPython; IPython.embed()
if Sub(interval3.a, interval6.b) > epsilon or Sub(interval6.a,
interval3.b) > epsilon:
log.debug("interval3 and interval6 do not intersect each other")
return ParabolicCurve()
# interval7 = interval3 \cap interval6
interval7 = iv.mpf(
[max(interval3.a, interval6.a),
min(interval3.b, interval6.b)])
if Sub(interval0.a, interval7.b) > epsilon or Sub(interval7.a,
interval0.b) > epsilon:
log.debug("interval0 and interval7 do not intersect each other")
return ParabolicCurve()
# interval8 = interval0 \cap interval7 : valid interval of t0 when considering all constraints (from a0 and a1)
interval8 = iv.mpf(
[max(interval0.a, interval7.a),
min(interval0.b, interval7.b)])
# import IPython; IPython.embed()
# We choose the value t0 (the duration of the first ramp) by selecting the mid point of the
# valid interval of t0.
t0 = _SolveForT0(A, B, newDuration, interval8)
if t0 is None:
# The fancy procedure fails. Now consider no optimization whatsoever.
# TODO: Figure out why solving fails.
t0 = mp.convert(interval8.mid) # select the midpoint
# return ParabolicCurve()
t1 = Sub(newDuration, t0)
a0 = Add(A, Mul(mp.fdiv(one, t0), B))
if (Abs(t1) < epsilon):
a1 = zero
else:
a1 = Add(A, Mul(mp.fdiv(one, Neg(t1)), B))
assert (Sub(Abs(a0), am) < epsilon
) # check if a0 is really below the bound
assert (Sub(Abs(a1), am) < epsilon
) # check if a1 is really below the bound
# import IPython; IPython.embed()
# Check if the velocity bound is violated
vp = Add(v0, Mul(a0, t0))
if Abs(vp) > vm:
vmnew = Mul(mp.sign(vp), vm)
D2 = Prod([
pointfive,
Sqr(Sub(vp, vmnew)),
Sub(mp.fdiv(one, a0), mp.fdiv(one, a1))
])
# print "D2",
# mp.nprint(D2, n=_prec)
# print "vmnew",
# mp.nprint(vmnew, n=_prec)
A2 = Sqr(Sub(vmnew, v0))
B2 = Neg(Sqr(Sub(vmnew, v1)))
t0trimmed = mp.fdiv(Sub(vmnew, v0), a0)
t1trimmed = mp.fdiv(Sub(v1, vmnew), a1)
C2 = Sum([
Mul(t0trimmed, Sub(vmnew, v0)),
Mul(t1trimmed, Sub(vmnew, v1)),
Mul(mp.mpf('-2'), D2)
])
log.debug("\nA2 = {0}; \nB2 = {1}; \nC2 = {2}; \nD2 = {3};".format(
mp.nstr(A2, n=_prec), mp.nstr(B2, n=_prec), mp.nstr(C2, n=_prec),
mp.nstr(D2, n=_prec)))
temp = Prod([A2, B2, B2])
initguess = mp.sign(temp) * (Abs(temp)**(1. / 3.))
root = mp.findroot(lambda x: Sub(Prod([x, x, x]), temp), x0=initguess)
# import IPython; IPython.embed()
log.debug("root = {0}".format(mp.nstr(root, n=_prec)))
a0new = mp.fdiv(Add(A2, root), C2)
if (Abs(a0new) > Add(am, epsilon)):
if FuzzyZero(Sub(Mul(C2, a0new), A2), epsilon):
# The computed a0new is exceeding the bound and its corresponding a1new is
# zero. Therefore, there is no other way to fix this. This is probably because the
# given newDuration is less than the minimum duration (x0, x1, v0, v1, vm, am) can
# get.
log.debug(
"abs(a0new) > am and a1new = 0; Cannot fix this case. This happens probably because the given newDuration is too short."
)
return ParabolicCurve()
a0new = Mul(mp.sign(a0new), am)
if (Abs(a0new) < epsilon):
a1new = mp.fdiv(B2, C2)
if (Abs(a1new) > Add(am, epsilon)):
# Similar to the case above
log.debug(
"a0new = 0 and abs(a1new) > am; Cannot fix this case. This happens probably because the given newDuration is too short."
)
return ParabolicCurve()
else:
if FuzzyZero(Sub(Mul(C2, a0new), A2), epsilon):
# import IPython; IPython.embed()
a1new = 0
else:
a1new = Mul(mp.fdiv(B2, C2),
Add(one, mp.fdiv(A2, Sub(Mul(C2, a0new), A2))))
if (Abs(a1new) > Add(am, epsilon)):
a1new = Mul(mp.sign(a1new), am)
a0new = Mul(mp.fdiv(A2, C2),
Add(one, mp.fdiv(B2, Sub(Mul(C2, a1new), B2))))
if (Abs(a0new) > Add(am, epsilon)) or (Abs(a1new) > Add(am, epsilon)):
log.warn("Cannot fix acceleration bounds violation")
return ParabolicCurve()
log.debug(
"\na0 = {0}; \na0new = {1}; \na1 = {2}; \na1new = {3};".format(
mp.nstr(a0, n=_prec), mp.nstr(a0new, n=_prec),
mp.nstr(a1, n=_prec), mp.nstr(a1new, n=_prec)))
if (Abs(a0new) < epsilon) and (Abs(a1new) < epsilon):
log.warn("Both accelerations are zero. Should we allow this case?")
return ParabolicCurve()
if (Abs(a0new) < epsilon):
# This is likely because v0 is at the velocity bound
t1new = mp.fdiv(Sub(v1, vmnew), a1new)
assert (t1new > 0)
ramp2 = Ramp(v0, a1new, t1new)
t0new = Sub(newDuration, t1new)
assert (t0new > 0)
ramp1 = Ramp(v0, zero, t0new, x0)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
elif (Abs(a1new) < epsilon):
t0new = mp.fdiv(Sub(vmnew, v0), a0new)
assert (t0new > 0)
ramp1 = Ramp(v0, a0new, t0new, x0)
t1new = Sub(newDuration, t0new)
assert (t1new > 0)
ramp2 = Ramp(ramp1.v1, zero, t1new)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
else:
# No problem with those new accelerations
# import IPython; IPython.embed()
t0new = mp.fdiv(Sub(vmnew, v0), a0new)
if (t0new < 0):
log.debug(
"t0new < 0. The given newDuration not achievable with the given bounds"
)
return ParabolicCurve()
t1new = mp.fdiv(Sub(v1, vmnew), a1new)
if (t1new < 0):
log.debug(
"t1new < 0. The given newDuration not achievable with the given bounds"
)
return ParabolicCurve()
if (Add(t0new, t1new) > newDuration):
# Final fix. Since we give more weight to acceleration bounds, we make the velocity
# bound saturated. Therefore, we set vp to vmnew.
# import IPython; IPython.embed()
if FuzzyZero(A, epsilon):
log.warn(
"(final fix) A is zero. Don't know how to fix this case"
)
return ParabolicCurve()
t0new = mp.fdiv(Sub(Sub(vmnew, v0), B), A)
if (t0new < 0):
log.debug("(final fix) t0new is negative")
return ParabolicCurve()
t1new = Sub(newDuration, t0new)
a0new = Add(A, Mul(mp.fdiv(one, t0new), B))
a1new = Add(A, Mul(mp.fdiv(one, Neg(t1new)), B))
ramp1 = Ramp(v0, a0new, t0new, x0)
ramp2 = Ramp(ramp1.v1, a1new, t1new)
newCurve = ParabolicCurve([ramp1, ramp2])
else:
ramp1 = Ramp(v0, a0new, t0new, x0)
ramp3 = Ramp(ramp1.v1, a1new, t1new)
ramp2 = Ramp(ramp1.v1, zero, Sub(newDuration,
Add(t0new, t1new)))
newCurve = ParabolicCurve([ramp1, ramp2, ramp3])
# import IPython; IPython.embed()
return newCurve
else:
ramp1 = Ramp(v0, a0, t0, x0)
ramp2 = Ramp(ramp1.v1, a1, t1)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
def Interpolate1DFixedDuration(x0, x1, v0, v1, newDuration, vm, am):
x0 = ConvertFloatToMPF(x0)
x1 = ConvertFloatToMPF(x1)
v0 = ConvertFloatToMPF(v0)
v1 = ConvertFloatToMPF(v1)
vm = ConvertFloatToMPF(vm)
am = ConvertFloatToMPF(am)
newDuration = ConvertFloatToMPF(newDuration)
log.debug("\nx0 = {0}; x1 = {1}; v0 = {2}; v1 = {3}; vm = {4}; am = {5}; newDuration = {6}".\
format(mp.nstr(x0, n=_prec), mp.nstr(x1, n=_prec), mp.nstr(v0, n=_prec), mp.nstr(v1, n=_prec),
mp.nstr(vm, n=_prec), mp.nstr(am, n=_prec), mp.nstr(newDuration, n=_prec)))
# Check inputs
assert(vm > zero)
assert(am > zero)
if (newDuration < -epsilon):
log.info("duration = {0} is negative".format(newDuration))
return ParabolicCurve()
if (newDuration <= epsilon):
# Check if this is a stationary trajectory
if (FuzzyEquals(x0, x1, epsilon) and FuzzyEquals(v0, v1, epsilon)):
log.info("stationary trajectory")
ramp0 = Ramp(v0, 0, 0, x0)
newCurve = ParabolicCurve(ramp0)
return newCurve
else:
log.info("newDuration is too short for any movement to be made")
return ParabolicCurve()
# Correct small discrepancies if any
if (v0 > vm):
if FuzzyEquals(v0, vm, epsilon):
v0 = vm
else:
log.info("v0 > vm: {0} > {1}".format(v0, vm))
return ParabolicCurve()
elif (v0 < -vm):
if FuzzyEquals(v0, -vm, epsilon):
v0 = -vm
else:
log.info("v0 < -vm: {0} < {1}".format(v0, -vm))
return ParabolicCurve()
if (v1 > vm):
if FuzzyEquals(v1, vm, epsilon):
v1 = vm
else:
log.info("v1 > vm: {0} > {1}".format(v1, vm))
return ParabolicCurve()
elif (v1 < -vm):
if FuzzyEquals(v1, -vm, epsilon):
v1 = -vm
else:
log.info("v1 < -vm: {0} < {1}".format(v1, -vm))
return ParabolicCurve()
d = Sub(x1, x0)
# First assume no velocity bound -> re-interpolated trajectory will have only two ramps.
# Solve for a0 and a1 (the acceleration of the first and the last ramps).
# a0 = A + B/t0
# a1 = A + B/(t - t0)
# where t is the (new) total duration, t0 is the (new) duration of the first ramp, and
# A = (v1 - v0)/t
# B = (2d/t) - (v0 + v1).
newDurInverse = mp.fdiv(one, newDuration)
A = Mul(Sub(v1, v0), newDurInverse)
B = Sub(Prod([mp.mpf('2'), d, newDurInverse]), Add(v0, v1))
interval0 = iv.mpf([zero, newDuration]) # initial interval for t0
# Now consider the interval(s) computed from a0's constraints
sum1 = Neg(Add(am, A))
sum2 = Sub(am, A)
C = mp.fdiv(B, sum1)
D = mp.fdiv(B, sum2)
log.debug("\nA = {0}; \nB = {1}; \nC = {2}; \nD = {3}; \nsum1 = {4}; \nsum2 = {5};".\
format(mp.nstr(A, n=_prec), mp.nstr(B, n=_prec), mp.nstr(C, n=_prec), mp.nstr(D, n=_prec),
mp.nstr(sum1, n=_prec), mp.nstr(sum2, n=_prec)))
if (sum1 > zero):
# This implied that the duration is too short
log.debug("the given duration ({0}) is too short.".format(newDuration))
return ParabolicCurve()
if (sum2 < zero):
# This implied that the duration is too short
log.debug("the given duration ({0}) is too short.".format(newDuration))
return ParabolicCurve()
if IsEqual(sum1, zero):
raise NotImplementedError # not yet considered
elif sum1 > epsilon:
log.debug("sum1 > 0. This implies that newDuration is too short.")
return ParabolicCurve()
else:
interval1 = iv.mpf([C, inf])
if IsEqual(sum2, zero):
raise NotImplementedError # not yet considered
elif sum2 > epsilon:
interval2 = iv.mpf([D, inf])
else:
log.debug("sum2 < 0. This implies that newDuration is too short.")
return ParabolicCurve()
if Sub(interval2.a, interval1.b) > epsilon or Sub(interval1.a, interval2.b) > epsilon:
# interval1 and interval2 do not intersect each other
return ParabolicCurve()
# interval3 = interval1 \cap interval2 : valid interval for t0 computed from a0's constraints
interval3 = iv.mpf([max(interval1.a, interval2.a), min(interval1.b, interval2.b)])
# Now consider the interval(s) computed from a1's constraints
if IsEqual(sum1, zero):
raise NotImplementedError # not yet considered
elif sum1 > epsilon:
log.debug("sum1 > 0. This implies that newDuration is too short.")
return ParabolicCurve()
else:
interval4 = iv.mpf([Neg(inf), Add(C, newDuration)])
if IsEqual(sum2, zero):
raise NotImplementedError # not yet considered
elif sum2 > epsilon:
interval5 = iv.mpf([Neg(inf), Add(D, newDuration)])
else:
log.debug("sum2 < 0. This implies that newDuration is too short.")
return ParabolicCurve()
if Sub(interval5.a, interval4.b) > epsilon or Sub(interval4.a, interval5.b) > epsilon:
log.debug("interval4 and interval5 do not intersect each other")
return ParabolicCurve()
# interval6 = interval4 \cap interval5 : valid interval for t0 computed from a1's constraints
interval6 = iv.mpf([max(interval4.a, interval5.a), min(interval4.b, interval5.b)])
# import IPython; IPython.embed()
if Sub(interval3.a, interval6.b) > epsilon or Sub(interval6.a, interval3.b) > epsilon:
log.debug("interval3 and interval6 do not intersect each other")
return ParabolicCurve()
# interval7 = interval3 \cap interval6
interval7 = iv.mpf([max(interval3.a, interval6.a), min(interval3.b, interval6.b)])
if Sub(interval0.a, interval7.b) > epsilon or Sub(interval7.a, interval0.b) > epsilon:
log.debug("interval0 and interval7 do not intersect each other")
return ParabolicCurve()
# interval8 = interval0 \cap interval7 : valid interval of t0 when considering all constraints (from a0 and a1)
interval8 = iv.mpf([max(interval0.a, interval7.a), min(interval0.b, interval7.b)])
# import IPython; IPython.embed()
# We choose the value t0 (the duration of the first ramp) by selecting the mid point of the
# valid interval of t0.
t0 = _SolveForT0(A, B, newDuration, interval8)
if t0 is None:
# The fancy procedure fails. Now consider no optimization whatsoever.
# TODO: Figure out why solving fails.
t0 = mp.convert(interval8.mid) # select the midpoint
# return ParabolicCurve()
t1 = Sub(newDuration, t0)
a0 = Add(A, Mul(mp.fdiv(one, t0), B))
if (Abs(t1) < epsilon):
a1 = zero
else:
a1 = Add(A, Mul(mp.fdiv(one, Neg(t1)), B))
assert(Sub(Abs(a0), am) < epsilon) # check if a0 is really below the bound
assert(Sub(Abs(a1), am) < epsilon) # check if a1 is really below the bound
# import IPython; IPython.embed()
# Check if the velocity bound is violated
vp = Add(v0, Mul(a0, t0))
if Abs(vp) > vm:
vmnew = Mul(mp.sign(vp), vm)
D2 = Prod([pointfive, Sqr(Sub(vp, vmnew)), Sub(mp.fdiv(one, a0), mp.fdiv(one, a1))])
# print("D2", end=' ')
# mp.nprint(D2, n=_prec)
# print("vmnew", end=' ')
# mp.nprint(vmnew, n=_prec)
A2 = Sqr(Sub(vmnew, v0))
B2 = Neg(Sqr(Sub(vmnew, v1)))
t0trimmed = mp.fdiv(Sub(vmnew, v0), a0)
t1trimmed = mp.fdiv(Sub(v1, vmnew), a1)
C2 = Sum([Mul(t0trimmed, Sub(vmnew, v0)), Mul(t1trimmed, Sub(vmnew, v1)), Mul(mp.mpf('-2'), D2)])
log.debug("\nA2 = {0}; \nB2 = {1}; \nC2 = {2}; \nD2 = {3};".format(mp.nstr(A2, n=_prec), mp.nstr(B2, n=_prec), mp.nstr(C2, n=_prec), mp.nstr(D2, n=_prec)))
temp = Prod([A2, B2, B2])
initguess = mp.sign(temp)*(Abs(temp)**(1./3.))
root = mp.findroot(lambda x: Sub(Prod([x, x, x]), temp), x0=initguess)
# import IPython; IPython.embed()
log.debug("root = {0}".format(mp.nstr(root, n=_prec)))
a0new = mp.fdiv(Add(A2, root), C2)
if (Abs(a0new) > Add(am, epsilon)):
if FuzzyZero(Sub(Mul(C2, a0new), A2), epsilon):
# The computed a0new is exceeding the bound and its corresponding a1new is
# zero. Therefore, there is no other way to fix this. This is probably because the
# given newDuration is less than the minimum duration (x0, x1, v0, v1, vm, am) can
# get.
log.debug("abs(a0new) > am and a1new = 0; Cannot fix this case. This happens probably because the given newDuration is too short.")
return ParabolicCurve()
a0new = Mul(mp.sign(a0new), am)
if (Abs(a0new) < epsilon):
a1new = mp.fdiv(B2, C2)
if (Abs(a1new) > Add(am, epsilon)):
# Similar to the case above
log.debug("a0new = 0 and abs(a1new) > am; Cannot fix this case. This happens probably because the given newDuration is too short.")
return ParabolicCurve()
else:
if FuzzyZero(Sub(Mul(C2, a0new), A2), epsilon):
# import IPython; IPython.embed()
a1new = 0
else:
a1new = Mul(mp.fdiv(B2, C2), Add(one, mp.fdiv(A2, Sub(Mul(C2, a0new), A2))))
if (Abs(a1new) > Add(am, epsilon)):
a1new = Mul(mp.sign(a1new), am)
a0new = Mul(mp.fdiv(A2, C2), Add(one, mp.fdiv(B2, Sub(Mul(C2, a1new), B2))))
if (Abs(a0new) > Add(am, epsilon)) or (Abs(a1new) > Add(am, epsilon)):
log.warn("Cannot fix acceleration bounds violation")
return ParabolicCurve()
log.debug("\na0 = {0}; \na0new = {1}; \na1 = {2}; \na1new = {3};".format(mp.nstr(a0, n=_prec), mp.nstr(a0new, n=_prec), mp.nstr(a1, n=_prec), mp.nstr(a1new, n=_prec)))
if (Abs(a0new) < epsilon) and (Abs(a1new) < epsilon):
log.warn("Both accelerations are zero. Should we allow this case?")
return ParabolicCurve()
if (Abs(a0new) < epsilon):
# This is likely because v0 is at the velocity bound
t1new = mp.fdiv(Sub(v1, vmnew), a1new)
assert(t1new > 0)
ramp2 = Ramp(v0, a1new, t1new)
t0new = Sub(newDuration, t1new)
assert(t0new > 0)
ramp1 = Ramp(v0, zero, t0new, x0)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
elif (Abs(a1new) < epsilon):
t0new = mp.fdiv(Sub(vmnew, v0), a0new)
assert(t0new > 0)
ramp1 = Ramp(v0, a0new, t0new, x0)
t1new = Sub(newDuration, t0new)
assert(t1new > 0)
ramp2 = Ramp(ramp1.v1, zero, t1new)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
else:
# No problem with those new accelerations
# import IPython; IPython.embed()
t0new = mp.fdiv(Sub(vmnew, v0), a0new)
if (t0new < 0):
log.debug("t0new < 0. The given newDuration not achievable with the given bounds")
return ParabolicCurve()
t1new = mp.fdiv(Sub(v1, vmnew), a1new)
if (t1new < 0):
log.debug("t1new < 0. The given newDuration not achievable with the given bounds")
return ParabolicCurve()
if (Add(t0new, t1new) > newDuration):
# Final fix. Since we give more weight to acceleration bounds, we make the velocity
# bound saturated. Therefore, we set vp to vmnew.
# import IPython; IPython.embed()
if FuzzyZero(A, epsilon):
log.warn("(final fix) A is zero. Don't know how to fix this case")
return ParabolicCurve()
t0new = mp.fdiv(Sub(Sub(vmnew, v0), B), A)
if (t0new < 0):
log.debug("(final fix) t0new is negative")
return ParabolicCurve()
t1new = Sub(newDuration, t0new)
a0new = Add(A, Mul(mp.fdiv(one, t0new), B))
a1new = Add(A, Mul(mp.fdiv(one, Neg(t1new)), B))
ramp1 = Ramp(v0, a0new, t0new, x0)
ramp2 = Ramp(ramp1.v1, a1new, t1new)
newCurve = ParabolicCurve([ramp1, ramp2])
else:
ramp1 = Ramp(v0, a0new, t0new, x0)
ramp3 = Ramp(ramp1.v1, a1new, t1new)
ramp2 = Ramp(ramp1.v1, zero, Sub(newDuration, Add(t0new , t1new)))
newCurve = ParabolicCurve([ramp1, ramp2, ramp3])
# import IPython; IPython.embed()
return newCurve
else:
ramp1 = Ramp(v0, a0, t0, x0)
ramp2 = Ramp(ramp1.v1, a1, t1)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
def _Stretch1D(curve, newDuration, vm, am):
log.debug("\nx0 = {0}; x1 = {1}; v0 = {2}; v1 = {3}; vm = {4}; am = {5}; prevDuration = {6}; newDuration = {7}".\
format(mp.nstr(curve.x0, n=_prec), mp.nstr(curve.EvalPos(curve.duration), n=_prec),
mp.nstr(curve.v0, n=_prec), mp.nstr(curve.EvalVel(curve.duration), n=_prec),
mp.nstr(vm, n=_prec), mp.nstr(am, n=_prec), mp.nstr(curve.duration, n=_prec),
mp.nstr(newDuration, n=_prec)))
# Check types
if type(newDuration) is not mp.mpf:
newDuration = mp.mpf("{:.15e}".format(newDuration))
if type(vm) is not mp.mpf:
vm = mp.mpf("{:.15e}".format(vm))
if type(am) is not mp.mpf:
am = mp.mpf("{:.15e}".format(am))
# Check inputs
# assert(newDuration > curve.duration)
assert(vm > zero)
assert(am > zero)
if (newDuration < -epsilon):
return ParabolicCurve()
if (newDuration <= epsilon):
# Check if this is a stationary trajectory
if (FuzzyEquals(curve.x0, curve.EvalPos(x1), epsilon) and FuzzyEquals(curve.v0, curve.v1, epsilon)):
ramp0 = Ramp(curve.v0, 0, 0, curve.x0)
newCurve = ParabolicCurve(ramp0)
return newCurve
else:
# newDuration is too short to any movement to be made
return ParabolicCurve()
v0 = curve[0].v0
v1 = curve[-1].v1
d = curve.d
# First assume no velocity bound -> re-interpolated trajectory will have only two ramps.
# Solve for a0 and a1 (the acceleration of the first and the last ramps).
# a0 = A + B/t0
# a1 = A + B/(t - t0)
# where t is the (new) total duration, t0 is the (new) duration of the first ramp, and
# A = (v1 - v0)/t
# B = (2d/t) - (v0 + v1).
newDurInverse = mp.fdiv(one, newDuration)
A = Mul(Sub(v1, v0), newDurInverse)
B = Sub(Prod([mp.mpf('2'), d, newDurInverse]), Add(v0, v1))
interval0 = iv.mpf([zero, newDuration]) # initial interval for t0
# Now consider the interval(s) computed from a0's constraints
sum1 = Neg(Add(am, A))
sum2 = Sub(am, A)
C = mp.fdiv(B, sum1)
D = mp.fdiv(B, sum2)
log.debug("\nA = {0}; \nB = {1}; \nC = {2}; \nD = {3}; \nsum1 = {4}; \nsum2 = {5};".\
format(mp.nstr(A, n=_prec), mp.nstr(B, n=_prec), mp.nstr(C, n=_prec), mp.nstr(D, n=_prec),
mp.nstr(sum1, n=_prec), mp.nstr(sum2, n=_prec)))
assert(not (sum1 > zero))
assert(not (sum2 < zero))
if IsEqual(sum1, zero):
raise NotImplementedError # not yet considered
elif sum1 > epsilon:
log.debug("sum1 > 0. This implies that newDuration is too short.")
return ParabolicCurve()
else:
interval1 = iv.mpf([C, inf])
if IsEqual(sum2, zero):
raise NotImplementedError # not yet considered
elif sum2 > epsilon:
interval2 = iv.mpf([D, inf])
else:
log.debug("sum2 < 0. This implies that newDuration is too short.")
return ParabolicCurve()
if Sub(interval2.a, interval1.b) > epsilon or Sub(interval1.a, interval2.b) > epsilon:
# interval1 and interval2 do not intersect each other
return ParabolicCurve()
# interval3 = interval1 \cap interval2 : valid interval for t0 computed from a0's constraints
interval3 = iv.mpf([max(interval1.a, interval2.a), min(interval1.b, interval2.b)])
# Now consider the interval(s) computed from a1's constraints
if IsEqual(sum1, zero):
raise NotImplementedError # not yet considered
elif sum1 > epsilon:
log.debug("sum1 > 0. This implies that newDuration is too short.")
return ParabolicCurve()
else:
interval4 = iv.mpf([Neg(inf), Add(C, newDuration)])
if IsEqual(sum2, zero):
raise NotImplementedError # not yet considered
elif sum2 > epsilon:
interval5 = iv.mpf([Neg(inf), Add(D, newDuration)])
else:
log.debug("sum2 < 0. This implies that newDuration is too short.")
return ParabolicCurve()
if Sub(interval5.a, interval4.b) > epsilon or Sub(interval4.a, interval5.b) > epsilon:
# interval4 and interval5 do not intersect each other
return ParabolicCurve()
# interval6 = interval4 \cap interval5 : valid interval for t0 computed from a1's constraints
interval6 = iv.mpf([max(interval4.a, interval5.a), min(interval4.b, interval5.b)])
# import IPython; IPython.embed()
if Sub(interval3.a, interval6.b) > epsilon or Sub(interval6.a, interval3.b) > epsilon:
# interval3 and interval6 do not intersect each other
return ParabolicCurve()
# interval7 = interval3 \cap interval6
interval7 = iv.mpf([max(interval3.a, interval6.a), min(interval3.b, interval6.b)])
if Sub(interval0.a, interval7.b) > epsilon or Sub(interval7.a, interval0.b) > epsilon:
# interval0 and interval7 do not intersect each other
return ParabolicCurve()
# interval8 = interval0 \cap interval7 : valid interval of t0 when considering all constraints (from a0 and a1)
interval8 = iv.mpf([max(interval0.a, interval7.a), min(interval0.b, interval7.b)])
# import IPython; IPython.embed()
# We choose the value t0 (the duration of the first ramp) by selecting the mid point of the
# valid interval of t0.
t0 = _SolveForT0(A, B, newDuration, interval8)
if t0 is None:
# The fancy procedure fails. Now consider no optimization whatsoever.
# TODO: Figure out why solving fails.
t0 = mp.convert(interval8.mid) # select the midpoint
# return ParabolicCurve()
t1 = Sub(newDuration, t0)
a0 = Add(A, Mul(mp.fdiv(one, t0), B))
if (Abs(t1) < epsilon):
a1 = zero
else:
a1 = Add(A, Mul(mp.fdiv(one, Neg(t1)), B))
assert(Sub(Abs(a0), am) < epsilon) # check if a0 is really below the bound
assert(Sub(Abs(a1), am) < epsilon) # check if a1 is really below the bound
# import IPython; IPython.embed()
# Check if the velocity bound is violated
vp = Add(v0, Mul(a0, t0))
if Abs(vp) > vm:
vmnew = Mul(mp.sign(vp), vm)
D2 = Prod([pointfive, Sqr(Sub(vp, vmnew)), Sub(mp.fdiv(one, a0), mp.fdiv(one, a1))])
# print "D2",
# mp.nprint(D2, n=_prec)
# print "vmnew",
# mp.nprint(vmnew, n=_prec)
A2 = Sqr(Sub(vmnew, v0))
B2 = Neg(Sqr(Sub(vmnew, v1)))
t0trimmed = mp.fdiv(Sub(vmnew, v0), a0)
t1trimmed = mp.fdiv(Sub(v1, vmnew), a1)
C2 = Sum([Mul(t0trimmed, Sub(vmnew, v0)), Mul(t1trimmed, Sub(vmnew, v1)), Mul(mp.mpf('-2'), D2)])
log.debug("\nA2 = {0}; \nB2 = {1}; \nC2 = {2}; \nD2 = {3};".format(mp.nstr(A2, n=_prec), mp.nstr(B2, n=_prec), mp.nstr(C2, n=_prec), mp.nstr(D2, n=_prec)))
temp = Prod([A2, B2, B2])
initguess = mp.sign(temp)*(Abs(temp)**(1./3.))
root = mp.findroot(lambda x: Sub(Prod([x, x, x]), temp), x0=initguess)
# import IPython; IPython.embed()
log.debug("root = {0}".format(mp.nstr(root, n=_prec)))
a0new = mp.fdiv(Add(A2, root), C2)
if (Abs(a0new) > Add(am, epsilon)):
a0new = Mul(mp.sign(a0new), am)
if (Abs(a0new) < epsilon):
a1new = mp.fdiv(B2, C2)
if (Abs(a1new) > Add(am, epsilon)):
log.warn("abs(a1new) > am; cannot fix this case")
# Cannot fix this case
return ParabolicCurve()
else:
if FuzzyZero(Sub(Mul(C2, a0new), A2), epsilon):
# import IPython; IPython.embed()
a1new = 0
else:
a1new = Mul(mp.fdiv(B2, C2), Add(one, mp.fdiv(A2, Sub(Mul(C2, a0new), A2))))
if (Abs(a1new) > Add(am, epsilon)):
a1new = Mul(mp.sign(a1new), am)
a0new = Mul(mp.fdiv(A2, C2), Add(one, mp.fdiv(B2, Sub(Mul(C2, a1new), B2))))
if (Abs(a0new) > Add(am, epsilon)) or (Abs(a1new) > Add(am, epsilon)):
log.warn("Cannot fix acceleration bounds violation")
return ParabolicCurve()
log.debug("\na0 = {0}; \na0new = {1}; \na1 = {2}; \na1new = {3};".format(mp.nstr(a0, n=_prec), mp.nstr(a0new, n=_prec), mp.nstr(a1, n=_prec), mp.nstr(a1new, n=_prec)))
if (Abs(a0new) < epsilon) and (Abs(a1new) < epsilon):
log.warn("Both accelerations are zero. Should we allow this case?")
return ParabolicCurve()
if (Abs(a0new) < epsilon):
# This is likely because v0 is at the velocity bound
t1new = mp.fdiv(Sub(v1, vmnew), a1new)
assert(t1new > 0)
ramp2 = Ramp(v0, a1new, t1new)
t0new = Sub(newDuration, t1new)
assert(t0new > 0)
ramp1 = Ramp(v0, zero, t0new, curve.x0)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
elif (Abs(a1new) < epsilon):
t0new = mp.fdiv(Sub(vmnew, v0), a0new)
assert(t0new > 0)
ramp1 = Ramp(v0, a0new, t0new, curve.x0)
t1new = Sub(newDuration, t0new)
assert(t1new > 0)
ramp2 = Ramp(ramp1.v1, zero, t1new)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
else:
# No problem with those new accelerations
# import IPython; IPython.embed()
t0new = mp.fdiv(Sub(vmnew, v0), a0new)
assert(t0new > 0)
t1new = mp.fdiv(Sub(v1, vmnew), a1new)
assert(t1new > 0)
if (Add(t0new, t1new) > newDuration):
# Final fix. Since we give more weight to acceleration bounds, we make the velocity
# bound saturated. Therefore, we set vp to vmnew.
# import IPython; IPython.embed()
t0new = mp.fdiv(Sub(Sub(vmnew, v0), B), A)
t1new = Sub(newDuration, t0new)
assert(t1new > zero)
a0new = Add(A, Mul(mp.fdiv(one, t0new), B))
a1new = Add(A, Mul(mp.fdiv(one, Neg(t1new)), B))
ramp1 = Ramp(v0, a0new, t0new, curve.x0)
ramp2 = Ramp(ramp1.v1, a1new, t1new)
newCurve = ParabolicCurve([ramp1, ramp2])
else:
# t0new = mp.fdiv(Sub(vmnew, v0), a0new)
# assert(t0new > 0)
ramp1 = Ramp(v0, a0new, t0new, curve.x0)
# t1new = mp.fdiv(Sub(v1, vmnew), a1new)
# assert(t1new > 0)
ramp3 = Ramp(ramp1.v1, a1new, t1new)
# print "a1",
# mp.nprint(a1, n=_prec)
# print "a1new",
# mp.nprint(a1new, n=_prec)
# print "t0trimmed",
# mp.nprint(t0trimmed, n=_prec)
# print "t0new",
# mp.nprint(t0new, n=_prec)
# print "t1trimmed",
# mp.nprint(t1trimmed, n=_prec)
# print "t1new",
# mp.nprint(t1new, n=_prec)
# print "T",
# mp.nprint(newDuration, n=_prec)
ramp2 = Ramp(ramp1.v1, zero, Sub(newDuration, Add(t0new , t1new)))
newCurve = ParabolicCurve([ramp1, ramp2, ramp3])
# import IPython; IPython.embed()
return newCurve
else:
ramp1 = Ramp(v0, a0, t0, curve.x0)
ramp2 = Ramp(ramp1.v1, a1, t1)
newCurve = ParabolicCurve([ramp1, ramp2])
return newCurve
def test_matrix_abs_max(self):
A = iv.matrix([[1, 2], [4, 5]]) + iv.mpf([-1, +1]) * mp.mpf(1e-10)
assert mp.matrix([[1, 2], [4, 5]]) + mp.mpf(1e-10) == matrix_abs_max(A)
assert mp.matrix([[1, 2], [4, 5]
]) + mp.mpf(1e-10) == matrix_abs_max(-A)
def set_state(self, state: Tuple[Tuple]):
self.state = tuple([iv.mpf([x[0], x[1]]) for x in state])
def test_membership(self):
assert iv.mpf([0.5, 0.5]) in iv.mpf([0.5, 0.5])
def beta(a,b):
return iv.mpf([-1,1])*iv.sqrt(lambda_s(a)*lambda_s(b))
def lambda_s(a):
return sum(iv.mpf([0,1])*max(ai.a**2,ai.b**2) for ai in a)
def run_analysis_continuized(self, name, outdir):
"""
Run Analysis with Continuization
see run_analysis()
WARNING: this will modify parameters in self.s (TODO fix that)
"""
print("")
print("====================")
print("Analysing System with Continuization: " + name)
assert self.s.immediate_ctrl
# FIXME: this will modify self.s (which is okay for how this function currently used, but may be annoying in the future)
self.s.continuize = True
# compute continuized dynamics
self.s.A_continuized_p_p = self.s.A_p + self.s.B_p @ self.s.B_d
self.s.A_continuized_p_c_tilde = self.s.B_p @ self.s.A_d
self.s.A_continuized_p_delta_p = self.s.B_p @ self.s.B_d
self.s.A_continuized_p_delta_c = self.s.A_continuized_p_c_tilde
self.s.A_continuized_c_tilde_p = 1/self.s.T * self.s.B_d
self.s.A_continuized_c_tilde_c_tilde = 1/self.s.T * (self.s.A_d - np.eye(self.s.n_d))
self.s.A_continuized_c_tilde_delta_p = self.s.A_continuized_c_tilde_p
self.s.A_continuized_c_tilde_delta_c = self.s.A_continuized_c_tilde_c_tilde
# build block matrices
# d/dt ([x_p; x_c_tilde]) = A_continuized_total @ [x_p; x_c_tilde] + B_continuized_total @ [delta_p; delta_c]
self.s.A_continuized_total = \
np.block([[self.s.A_continuized_p_p, self.s.A_continuized_p_c_tilde],
[self.s.A_continuized_c_tilde_p, self.s.A_continuized_c_tilde_c_tilde]])
self.s.B_continuized_total = \
np.block([[self.s.A_continuized_p_delta_p, self.s.A_continuized_p_delta_c],
[self.s.A_continuized_c_tilde_delta_p, self.s.A_continuized_c_tilde_delta_c]])
print("Eigenvalues of continuized system without delta:")
print(scipy.linalg.eig(self.s.A_continuized_total)[0])
self.s.spaceex_iterations = 1
# Note: SpaceEx uses unsound numerics (at least in the default settings).
# Epsilon here does not strictly guarantee soundness from a mathematical point of view (although it makes engineers sleep well).
eps = iv.mpf([-1e-10, 1e-10])
# \underline{\Delta}
delta_pc_assumed_without_eps = IntervalMatrix.zeros(self.s.n_p + self.s.n_d, 1)
i = 0
final_run = False
while True:
i += 1
print(f"\n\n\n# Iteration {i}")
if final_run:
print("(Final run - bound was already proven, now we generate all the plots and results)")
# run analysis
# NOTE: self.results now refers to continuized results
# \Delta = \underline{\Delta} + B_\epsilon (overapproximated as interval)
delta_pc_assumed = delta_pc_assumed_without_eps + eps
print("assumed bound for delta_p, delta_c:")
print(delta_pc_assumed)
self.s.delta_p_assumed = delta_pc_assumed[0:self.s.n_p]
self.s.delta_c_assumed = delta_pc_assumed[self.s.n_p:(self.s.n_p + self.s.n_d)]
self.run_analysis(name, outdir, extra_plots=final_run, print_header=False)
if self.results['spaceex_hypy']['code'] != hypy.Engine.SUCCESS:
print("SpaceEx failed. Giving up.")
self.results['continuization'] = 'fail (SpaceEx errored)'
return
# compute delta-bounds
x_pc_intv = IntervalMatrix.from_bounds(lower_upper=spaceex_bounds_to_array(self.results['spaceex_hypy']['output']['variables'])[0:self.s.n_p + self.s.n_d])
delta_pc_actual = iv.mpf([0, 1]) * self.s.T * (-1) * (IntervalMatrix.convert(self.s.A_continuized_total) @ (x_pc_intv + eps) + IntervalMatrix.convert(self.s.B_continuized_total) @ delta_pc_assumed)
print("resulting actual bound for delta_p, delta_c:")
print(delta_pc_actual)
# check delta-bounds
if final_run:
# this was the last run - see below for conditions
assert all([(delta_pc_actual[i]) in delta_pc_assumed_without_eps[i] for i in range(len(delta_pc_assumed))])
print("successfully analyzed")
self.results['continuization'] = 'success'
return
elif all([(delta_pc_actual[i]) in delta_pc_assumed_without_eps[i] for i in range(len(delta_pc_assumed))]):
print("assumed interval is correct. Done. Re-running analysis with final bounds.")
delta_pc_assumed_without_eps = delta_pc_actual
final_run = True
elif i < 5:
print("assumed interval for delta_p and/or delta_c is to small. Enlarging.")
delta_pc_assumed_without_eps = delta_pc_actual * iv.mpf([0, 2])
else:
print("Iteration did not converge. System is possibly unstable or otherwise unsuitable for Continuization.")
self.results['continuization'] = 'fail (no convergence)'
return
