def LCEB(fitness_fn, ngens, popsize, st_maxdepth):
"""Linear combination of evolved bases. There is a single
individual composed of many randomly-generated trees. At each
step, we take them as bases for a GLM, fit, and look at their
coefficients. Any tree which has a small coefficient is not
helping much: replace it with a new randomly-generated tree, and
repeat.
FIXME problem: it rewards trees which require huge coefficients,
ie hardly do anything."""
y = fitness_fn.train_y
# make initial population
pop = [gp.grow(st_maxdepth, random) for i in range(popsize)]
for gen in xrange(ngens):
X = None
for ind in pop:
fit, col = fitness_fn.get_semantics(gp.make_fn(ind))
if (fit != sys.maxint
and all(np.isfinite(col))):
pass
else:
print("Omitting a column")
col = np.zeros(len(y))
if X is None:
X = col
else:
X = np.c_[X, col]
print("X")
print(X.shape)
print(X)
model = LinearRegression()
model.fit(X, y)
coefs = model.coef_
output = model.predict(X)
rmse = fitness_fn.rmse(y, output)
print("rmse", rmse)
print("coefs", coefs)
worst_idx, worst_val = argabsmin(coefs)
print("worst tree")
print(pop[worst_idx])
pop[worst_idx] = gp.grow(st_maxdepth, random)
def test():
t, dep, nnodes = bubble_down(random.randint(5, 19), random)
s, dep, nnodes = bubble_down(random.randint(10, 15), random)
print(t)
print(s)
tN = gp.make_zssNode_from_tuple(t)
sN = gp.make_zssNode_from_tuple(s)
print(tN)
print(sN)
print(zss.compare.distance(tN, sN))
print(zss.compare.distance(tN, tN))
print(zss.compare.distance(sN, sN))
print("---")
cases = [(x, y) for x in range(4) for y in range(4)]
print(cases)
cases = zip(*cases)
print(cases)
tf = gp.make_fn(t)
print(tf(cases))
def generate_bubble_down_tree_and_fn(rng):
t, d, n = bubble_down(30, rng)
return t, gp.make_fn(t)
def generate_grow_tree_and_fn_maxd(rng, max_depth=3):
t = gp.grow(max_depth, rng)
return t, gp.make_fn(t)
def generate_bubble_down_tree_and_fn_minn_maxn(minn, maxn, rng):
t, d, n = bubble_down_minn_maxn(minn, maxn, rng)
return t, gp.make_fn(t)