def sobol_adaptive(f, dim, S_cutoff=0.95, max_samples=200, max_iters=10,
max_level=10, max_k=10,
fval=None, plotting=False, labels=None):
"""
Sobol-based dimension-adaptive sparse grids.
f - function to sample
dim - number of input variables/dimensions
S_cutoff - Sobol index cutoff, adapt dimensions with Sobol indices
adding up to the cutoff
max_samples - termination criteria, maximum allowed samples of f
max_iters - termination criteria, maximum adaptation iterations
max_level - maximum level allowed in any single variable, ie. don't allow
very-high resolution in any direction. Enforce
max(multiindex) <= max_level
max_k - enforce |multiindex|_1 <= dim + max_k - 1, ie. constrain
to simplex-rule of level max_k
fval - if samples of f already exist at sparse-grid nodes, pass
dictionary containing values - these will be used first
The iteration will also terminate if the grid is unchanged after an
iteration.
"""
# Initialize with simplex level 2, ie.
# one-factor, 3 points in each direction
K = IndexSet(dim=dim, level=2, fill='simplex')
quad = QuadratureCC()
sp = SparseGrid(dim, quad, indexset=K)
iter = 1
fval = {} if fval is None else fval # Dictionary of function values
# Main adaptation loop
while sp.n_nodes() <= max_samples and iter <= max_iters:
# Sampling call, don't recompute already
# known values
sp.sample_fn(f, fargs=(), fval=fval)
print('Iter', iter, '='*100)
print('sp.n_nodes() =', sp.n_nodes())
if plotting:
sp.plot(outfile='tmp/sobol_adapt_iter%d.pdf'%iter, labels=labels)
# 2. Compute Sobol indices, up to maximum
# interaction defined by the current K
D, mu, var = sp.compute_sobol_variances(fval,
cardinality=K.max_interaction(),
levelrefine=2)
del D[()] # Remove variance (==var)
print('# %6d %12.6e %12.6e' % (sp.n_nodes(), mu, var))
### ------------------------------------------- RESULT <==
# 3. Interaction selection
# Sort according to variance large->small
print('D =', D)
Dsort = sorted(D.iteritems(), key=lambda (k,v): -v)
print('Dsort =', Dsort)
print('var =', var)
sobol_total,i,U = 0.,0,set([]) # Select most important interactions
while sobol_total < S_cutoff and i < len(Dsort):
sobol_total += Dsort[i][1] / var
U |= set([Dsort[i][0]])
i += 1
print('U =', U)
# 4. Interaction augmentation
# Find set of potential *new*
# interactions present in active set
A = K.activeset()
potential_interactions = set([interaction(a) for a in A]) - \
K.interactions()
print('A = ', A)
print('potential_interactions =', potential_interactions)
# Select potential new interactions
# satisfying
Uplus = set([])
for interac in potential_interactions:
all_subsets = set([])
for r in range(1, len(interac)):
all_subsets |= set(itertools.combinations(interac, r))
if all_subsets <= U:
Uplus |= set([interac])
U |= Uplus
print('Uplus =', Uplus)
print('new U =', U)
# 5. Indexset extension - new sparse grid
unchanged = True
for a in A:
if np.sum(a) > max_k+dim-1: # Enforce simplex-constraint on indexset
continue
if np.max(a) > max_level: # Enforce maximum level constraint
continue
if interaction(a) in U:
unchanged = False
K.I |= set([a])
if unchanged:
print('No new multi-indices added to index-set, terminate adapation')
break
print('K.I =', K.I)
sp.set_indexset(K)
iter += 1
def gerstnerandgriebel_adaptive(f, dim, max_samples=200, max_iters=10,
min_error=1.e-16, max_level=10, max_k=10,
fval=None, plotting=False, labels=None):
"""
Gerstner+Griebel style dimension-adaptive sparse grids.
f - function to sample
dim - number of input variables/dimensions
max_samples - termination criteria, maximum allowed samples of f
max_iters - termination criteria, maximum adaptation iterations
max_level - maximum level allowed in any single variable, ie. don't allow
very-high resolution in any direction. Enforce
max(multiindex) <= max_level
max_k - enforce |multiindex|_1 <= dim + max_k - 1, ie. constrain
to simplex-rule of level max_k
fval - if samples of f already exist at sparse-grid nodes, pass
dictionary containing values - these will be used first
"""
# Initialize with simplex level 1
K = IndexSet(dim=dim, level=1, fill='simplex')
quad = QuadratureCC()
sp = SparseGrid(dim, quad, indexset=K)
iter, eta = 1, 1e100
fval = {} if fval is None else fval # Dictionary of function values
# Main adaptation loop
while sp.n_nodes() <= max_samples and iter <= max_iters and eta > min_error:
# Sampling call, don't recompute already
# known values
sp.sample_fn(f, fargs=(), fval=fval)
print('Iter', iter, '='*100)
print('sp.n_nodes() =', sp.n_nodes())
if plotting:
sp.plot(outfile='tmp/GandG_adapt_iter%d.png'%iter, labels=labels)
r = sp.integrate(fval)
print('r =', r)
# For each member of the active set,
# compute the difference between the
# objective, with and without that member
A = K.activeset()
g = {}
for a in A:
if np.sum(a) > max_k+dim-1: # Enforce simplex-constraint on indexset
continue
if np.max(a) > max_level: # Enforce maximum level constraint
continue
Kmod = copy.deepcopy(K)
Kmod.I |= set([a])
sp.set_indexset(Kmod)
sp.sample_fn(f, fargs=(), fval=fval)
rmod = sp.integrate(fval)
g[a] = abs(rmod - r)
if len(g) == 0:
print('No new multi-indices added to index-set, terminate adapation')
break
a_adapt = max(g, key=g.get)
eta = sum(g.values())
print('eta =',eta)
K.I |= set([a_adapt])
sp.set_indexset(K)
iter += 1
return r
