def dissimilarity(self, other, c=(5, 5, 5, 1, 5, 1)):
"""
The distance/dissimilarity of two genomes, similar to NEAT
dist = (c0*S + c1*D + c2*E)/N + c3*T + c4*K + c5*X
where
S sum of difference in same genes
D number of disjoint genes
E number of excess genes
N length of larger gene
T difference in optimizer + hyperparameters
K mean of difference in same nodes
X difference in trained epochs
"""
genes_1, genes_2 = map(lambda x: x.genes_by_id, [self, other])
ids_1, ids_2 = map(lambda x: set(x.keys()), [genes_1, genes_2])
nodes_1, nodes_2 = map(lambda x: x.nodes_by_id, [self, other])
node_ids = set(nodes_1.keys()) & set(nodes_2.keys())
N = 1 # TODO: max(len(ids_1), len(ids_2))
excess_start = max(ids_1 | ids_2) + 1
S = sum([
genes_1[_id].dissimilarity(genes_2[_id]) for _id in ids_1 & ids_2
])
D = len([_id for _id in ids_1 ^ ids_2 if _id < excess_start])
E = len(ids_1 ^ ids_2) - D
T = self.optimizer.dissimilarity(other.optimizer)
K = sum([nodes_1[_id].dissimilarity(nodes_2[_id])
for _id in node_ids]) / len(node_ids)
X = limited_growth(np.abs(self.trained - other.trained), 1, 10)
return (c[0] * S + c[1] * D +
c[2] * E) / N + c[3] * T + c[4] * K + c[5] * X
def dissimilarity(self, other):
if not isinstance(other, self.__class__):
return 1
dist = np.array([
self.size_change - other.size_change,
self.activation != other.activation
])
importance = np.array([0.6, 0.4])
relevance = np.array([80, 0.01])
return np.sum(limited_growth(np.abs(dist), importance, relevance))
def dissimilarity(self, other):
if not isinstance(other, self.__class__):
return 1
dist = np.array([
self.height - other.height, self.width - other.width,
self.stride - other.stride, self.padding - other.padding,
self.pooling != other.pooling
])
importance = np.array([0.2, 0.2, 0.1, 0.1, 0.4])
relevance = np.array([5, 5, 3, 3, 0.01])
return np.sum(limited_growth(np.abs(dist), importance, relevance))
def dissimilarity(self, other):
if type(other) != ADAMGene:
return 1
dist = np.array([
self.log_learning_rate - other.log_learning_rate,
self.log_weight_decay - other.log_weight_decay
])
importance = np.array([0.8, 0.2])
relevance = np.array([5, 3])
return np.sum(limited_growth(np.abs(dist), importance, relevance))
def dissimilarity(self, other):
if not isinstance(other, self.__class__):
return 1
dist = np.array([
self.height - other.height, self.width - other.width,
self.stride - other.stride, self.padding - other.padding,
self.depth_size_change - other.depth_size_change,
self.depth_mult - other.depth_mult
])
importance = np.array([0.2, 0.2, 0.1, 0.05, 0.1, 0.35])
relevance = np.array([5, 5, 3, 3, 5, 8])
return np.sum(limited_growth(np.abs(dist), importance, relevance))