def __init__(self, Sfeats, R=5.0):
"""
@param R radius of a neighborhood.
..note: R should be automatically infered, perhaps on a goal/effect basis.
Right now, we don't have a good heuristic to determine R. So it is
set arbitrarily.
"""
self.dim = len(Sfeats)
self.Sfeats = Sfeats
self.size = 0
self.R = R
self.goals = Databag(self.dim) # goal pursued
self.effects = Databag(self.dim) # effects observed
self.effects_interest = []
self.goals_interest = Meshgrid(len(Sfeats), 3, R)
# this is a meshgrid, as picking proportionnaly out of a sum of gaussian
# is intractable.
self.competence = []
def __init__(self, Sfeats, R = 5.0):
"""
@param R radius of a neighborhood.
..note: R should be automatically infered, perhaps on a goal/effect basis.
Right now, we don't have a good heuristic to determine R. So it is
set arbitrarily.
"""
self.dim = len(Sfeats)
self.Sfeats = Sfeats
self.size = 0
self.R = R
self.goals = Databag(self.dim) # goal pursued
self.effects = Databag(self.dim) # effects observed
self.effects_interest = []
self.goals_interest = Meshgrid(len(Sfeats), 3, R)
# this is a meshgrid, as picking proportionnaly out of a sum of gaussian
# is intractable.
self.competence = []
class LichenExplorer(GoalExplorer):
"""Lichen Explorer
Implementation of the Lichen algorithm for picking new goals
with a dual mode : effect and goal.
"""
def __init__(self, Sfeats, R=5.0):
"""
@param R radius of a neighborhood.
..note: R should be automatically infered, perhaps on a goal/effect basis.
Right now, we don't have a good heuristic to determine R. So it is
set arbitrarily.
"""
self.dim = len(Sfeats)
self.Sfeats = Sfeats
self.size = 0
self.R = R
self.goals = Databag(self.dim) # goal pursued
self.effects = Databag(self.dim) # effects observed
self.effects_interest = []
self.goals_interest = Meshgrid(len(Sfeats), 3, R)
# this is a meshgrid, as picking proportionnaly out of a sum of gaussian
# is intractable.
self.competence = []
def add_effect(self, effect, goal=None):
self.size += 1
effect_bare = effect[list(self.Sfeats)]
goal_bare = goal[list(self.Sfeats)] if goal is not None else None
if goal is not None:
self._update_goal_interest(goal_bare, effect_bare)
self._update_effect_interest(goal_bare, effect_bare)
def _update_goal_interest(self, goal_bare, effect_bare):
# 1. compute competence
c = competence.competence(goal_bare, effect_bare)
if len(self.goals) > 0:
# 2. find neighborhood
dists, nhood_idx = self.goals.nn(goal_bare, k=10)
#dists, nhood_idx = zip(*[(d, i) for d, i in zip(dists, nhood_idx) if d < 2*self.R])
# 3. compute interest
past_c = [self.competence[i] for i in sorted(nhood_idx)] + [c]
inte = max(0.0, interest.cealled_derivative(past_c))
# 4. assign interest
self.goals_interest.add_p(goal_bare, inte)
# for d, i in zip(dists, nhood_idx):
# w = math.exp(-d/(2*self.R*self.R))
# self.interest_g[i] = self.interest_g[i]*(1-w) + w*inte#inte
# #self.interest_g.append(inte)
# self.interest_g.append(inte)
else:
pass
# self.interest_g.append(0.01) # HACK HACK HACK -> Why not +inf ?
self.goals.add(goal_bare)
self.competence.append(c)
def _set_effect_interest(self, effect_idx):
effect = self.effects.get(effect_idx)
dists = self.goals.nn(effect, k=5)[0]
dists = [d for d in dists if d < float('inf')]
d = sum(dists) / len(dists)
# 2 compute interest
if d <= self.R:
self.effects_interest[effect_idx] = 0.0
else:
self.effects_interest[effect_idx] = d - self.R
def _update_effect_interest(self, goal_bare, effect_bare):
"""Update the interest of a new effect"""
self.effects.add(effect_bare)
# 1 find nearest goals
if len(self.goals) == 0:
self.effects_interest.append(
10 * self.R) # next motor or sensory babbling
# will be around there.
else:
self.effects_interest.append(None)
self._set_effect_interest(len(self.effects_interest) - 1)
# 3 override effects'interest in goal neighborhood
if goal_bare is not None:
dist, nhood_idx = self.effects.nn(goal_bare, k=20,
radius=self.R) # TODO k = 100 ?
for d, idx in zip(dist, nhood_idx):
if d < float('+inf'):
self._set_effect_interest(idx)
def next_goal(self):
# choose a past goal
if 100 * random.random() < goal_freq:
return self._mode_goal()
else:
return self._mode_effect()
def _mode_effect(self):
# choose an past effect
assert len(
self.effects_interest
) > 0, "Error : No motor babbling was done to bootstrap the space."
idx = toolbox.fun.roulette_wheel(self.effects_interest)
ref_effect = self.effects.get(idx)
gen_goal = ref_effect + self.R * np.array(
[random.uniform(-1, 1) for _ in self.Sfeats])
return pandas.Series(gen_goal, index=self.Sfeats)
def _mode_goal(self):
# choose a past goal
if len(self.goals_interest) == 0:
assert len(self.effects_interest) > 0
return self._mode_effect()
else:
areas, inte = zip(
*[(a, i) for a, i in self.goals_interest.nodes.iteritems()])
idx = toolbox.fun.roulette_wheel(inte)
ref_goal = self.goals_interest.res * np.array(areas[idx])
# generate a random goal
i = 0
while (i < 100):
i += 1
gen_goal = ref_goal + spread * self.R * np.array(
[random.uniform(-1, 1) for _ in self.Sfeats])
nn_effect = self.effects.get(
self.effects.nn(gen_goal, 1)[1][0])
if True or toolbox.dist(gen_goal, nn_effect) < spread * self.R:
return pandas.Series(gen_goal, index=self.Sfeats)
raise ValueError("Too much tries")
class LichenExplorer(GoalExplorer):
"""Lichen Explorer
Implementation of the Lichen algorithm for picking new goals
with a dual mode : effect and goal.
"""
def __init__(self, Sfeats, R = 5.0):
"""
@param R radius of a neighborhood.
..note: R should be automatically infered, perhaps on a goal/effect basis.
Right now, we don't have a good heuristic to determine R. So it is
set arbitrarily.
"""
self.dim = len(Sfeats)
self.Sfeats = Sfeats
self.size = 0
self.R = R
self.goals = Databag(self.dim) # goal pursued
self.effects = Databag(self.dim) # effects observed
self.effects_interest = []
self.goals_interest = Meshgrid(len(Sfeats), 3, R)
# this is a meshgrid, as picking proportionnaly out of a sum of gaussian
# is intractable.
self.competence = []
def add_effect(self, effect, goal = None):
self.size += 1
effect_bare = effect[list(self.Sfeats)]
goal_bare = goal[list(self.Sfeats)] if goal is not None else None
if goal is not None:
self._update_goal_interest(goal_bare, effect_bare)
self._update_effect_interest(goal_bare, effect_bare)
def _update_goal_interest(self, goal_bare, effect_bare):
# 1. compute competence
c = competence.competence(goal_bare, effect_bare)
if len(self.goals) > 0:
# 2. find neighborhood
dists, nhood_idx = self.goals.nn(goal_bare, k = 10)
#dists, nhood_idx = zip(*[(d, i) for d, i in zip(dists, nhood_idx) if d < 2*self.R])
# 3. compute interest
past_c = [self.competence[i] for i in sorted(nhood_idx)] + [c]
inte = max(0.0, interest.cealled_derivative(past_c))
# 4. assign interest
self.goals_interest.add_p(goal_bare, inte)
# for d, i in zip(dists, nhood_idx):
# w = math.exp(-d/(2*self.R*self.R))
# self.interest_g[i] = self.interest_g[i]*(1-w) + w*inte#inte
# #self.interest_g.append(inte)
# self.interest_g.append(inte)
else:
pass
# self.interest_g.append(0.01) # HACK HACK HACK -> Why not +inf ?
self.goals.add(goal_bare)
self.competence.append(c)
def _set_effect_interest(self, effect_idx):
effect = self.effects.get(effect_idx)
dists = self.goals.nn(effect, k = 5)[0]
dists = [d for d in dists if d < float('inf')]
d = sum(dists)/len(dists)
# 2 compute interest
if d <= self.R:
self.effects_interest[effect_idx] = 0.0
else:
self.effects_interest[effect_idx] = d-self.R
def _update_effect_interest(self, goal_bare, effect_bare):
"""Update the interest of a new effect"""
self.effects.add(effect_bare)
# 1 find nearest goals
if len(self.goals) == 0:
self.effects_interest.append(10*self.R) # next motor or sensory babbling
# will be around there.
else:
self.effects_interest.append(None)
self._set_effect_interest(len(self.effects_interest) - 1)
# 3 override effects'interest in goal neighborhood
if goal_bare is not None:
dist, nhood_idx = self.effects.nn(goal_bare, k = 20, radius = self.R) # TODO k = 100 ?
for d, idx in zip(dist, nhood_idx):
if d < float('+inf'):
self._set_effect_interest(idx)
def next_goal(self):
# choose a past goal
if 100*random.random() < goal_freq:
return self._mode_goal()
else:
return self._mode_effect()
def _mode_effect(self):
# choose an past effect
assert len(self.effects_interest) > 0, "Error : No motor babbling was done to bootstrap the space."
idx = toolbox.fun.roulette_wheel(self.effects_interest)
ref_effect = self.effects.get(idx)
gen_goal = ref_effect + self.R*np.array([random.uniform(-1, 1) for _ in self.Sfeats])
return pandas.Series(gen_goal, index = self.Sfeats)
def _mode_goal(self):
# choose a past goal
if len(self.goals_interest) == 0:
assert len(self.effects_interest) > 0
return self._mode_effect()
else:
areas, inte = zip(*[(a, i) for a, i in self.goals_interest.nodes.iteritems()])
idx = toolbox.fun.roulette_wheel(inte)
ref_goal = self.goals_interest.res * np.array(areas[idx])
# generate a random goal
i = 0
while (i < 100):
i += 1
gen_goal = ref_goal + spread*self.R*np.array([random.uniform(-1, 1) for _ in self.Sfeats])
nn_effect = self.effects.get(self.effects.nn(gen_goal, 1)[1][0])
if True or toolbox.dist(gen_goal, nn_effect) < spread*self.R:
return pandas.Series(gen_goal, index = self.Sfeats)
raise ValueError("Too much tries")
