Skip to content

Commit

Permalink
Apply PR feedback
Browse files Browse the repository at this point in the history
  • Loading branch information
@HenrikMettler
HenrikMettler committed Sep 4, 2020
1 parent daf0a65 commit 50f887c
Show file tree
Hide file tree
Showing 4 changed files with 72 additions and 127 deletions.
115 changes: 59 additions & 56 deletions cgp/genome.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -138,7 +138,7 @@ def __repr__(self) -> str:
return s

def determine_permissible_values_per_gene(self, gene_idx: int) -> np.ndarray:
region_idx = gene_idx // self._length_per_region
region_idx = self._get_region_idx(gene_idx)

if self._is_input_region(region_idx):
return self._determine_permissible_values_input_region(gene_idx)
Expand Down Expand Up @@ -273,16 +273,10 @@ def reorder(self, rng: np.random.RandomState) -> None:
Shuffle node ordering of internal (hidden) nodes in genome without changing node behavior.
(Goldman 2015, DOI: 10.1109/TEVC.2014.2324539)
From a set of nodes with no dependencies (addable_nodes)
a (pseudo-) random node is placed at the first internal position.
After updating the set of addable nodes,
a (pseudo-) random node for the next position is picked.
This procedure is repeated until all nodes are placed.
Gene values are updated to correspond to the new address of the corresponding node.
Invalid gene values (from unused arities of eg. constant operator nodes)
are replaced by (pseudo-) random picks of permissible values for that position.
During reordering, inactive genes, e.g., input genes of nodes with arity zero, are not taken
into account and can hence have invalid values after reordering.
These invalid values are replaced by random values
for the respective gene after reordering.
Parameters
----------
Expand All @@ -300,38 +294,38 @@ def reorder(self, rng: np.random.RandomState) -> None:
Warning("levels_back must be equal n_columns, reorder is not applied")
return

dna_for_reorder = self._dna.copy()
dna = self._dna.copy()

dependencies = self._determine_dependencies()
node_dependencies = self._determine_node_dependencies()

addable_nodes = set(int(idx_str) for idx_str, item in dependencies.items() if not item)
addable_nodes = self._get_addable_nodes(node_dependencies)

node_idx = self._n_inputs # First position to be placed is after inputs
used_idx_list: List = []
used_node_indices: List[int] = []

while len(addable_nodes) > 0:
# pick one of the addable nodes
current_idx = rng.choice(list(addable_nodes))

dna_for_reorder = self._exchange_dna_segments(
dna_for_reorder, idx_old_location=current_idx, idx_new_location=node_idx
current_node_idx = rng.choice(list(addable_nodes))

dna = self._copy_dna_segment(
dna, idx_old_location=current_node_idx, idx_new_location=node_idx
)

for _, current_dependencies in dependencies.items():
current_dependencies.discard(current_idx)
for dependencies in node_dependencies.values():
dependencies.discard(current_node_idx)

addable_nodes = self._update_addable_nodes(
dependencies, used_idx_list, idx_used=current_idx
[addable_nodes, used_node_indices] = self._update_addable_nodes(
node_dependencies, used_node_indices, current_idx=current_node_idx
)
node_idx += 1

self._update_gene_values(dna_for_reorder, used_idx_list)
self._replace_invalid_gene_values(dna_for_reorder, rng)
self._update_gene_values(dna, used_node_indices)
self._replace_invalid_gene_values(dna, rng)

self.dna = dna_for_reorder
self.dna = dna

def _exchange_dna_segments(
self, dna_new: List[int], idx_old_location: int, idx_new_location: int
def _copy_dna_segment(
self, dna: List[int], idx_old_location: int, idx_new_location: int
) -> List[int]:
""" Moves a nodes dna from its old node location to a new location. """

Expand All @@ -340,55 +334,67 @@ def _exchange_dna_segments(
* self._length_per_region : (idx_old_location + 1)
* self._length_per_region
]
dna_new[
dna[
idx_new_location
* self._length_per_region : (idx_new_location + 1)
* self._length_per_region
] = node_dna

return dna_new
return dna

def _update_addable_nodes(
self, dependencies: Dict, used_idx_list: List[int], idx_used: int
) -> set:
self, node_dependencies: Dict, used_node_indices: List[int], current_idx: int
) -> [set, List[int]]:
""" Update the set of addable nodes,
by reevaluating which nodes haves no dependencies
and remove nodes which were already placed from the list.
"""
addable_nodes = set(int(idx_str) for idx_str, item in dependencies.items() if not item)
used_idx_list.append(idx_used)
addable_nodes = addable_nodes.difference(used_idx_list)
addable_nodes = self._get_addable_nodes(node_dependencies)
used_node_indices.append(current_idx)
addable_nodes = addable_nodes.difference(used_node_indices)

return addable_nodes
return addable_nodes, used_node_indices

def _update_gene_values(self, dna_for_reorder: List[int], used_idx_list: List[int]) -> None:
def _update_gene_values(self, dna: List[int], used_node_indices: List[int]) -> None:
""" Update gene values to correspond to the new position of the corresponding node"""
for gene_idx, gene_value in enumerate(dna_for_reorder):
region_idx = gene_idx // self._length_per_region
for gene_idx, gene_value in enumerate(dna):
region_idx = self._get_region_idx(gene_idx)
if self._is_hidden_input_gene(gene_idx, region_idx) or self._is_output_input_gene(
gene_idx
):
try:
gene_value = self._n_inputs + used_idx_list.index(gene_value)
except ValueError: # means the input is from an input node -> do nothing
pass
dna_for_reorder[gene_idx] = gene_value
if gene_value >= self._n_inputs:
gene_value = self._n_inputs + used_node_indices.index(gene_value)
dna[gene_idx] = gene_value

def _replace_invalid_gene_values(
self, dna_for_reorder: List[int], rng: np.random.RandomState
) -> None:
""" Replace gene values of unused arities
by (pseudo-) random picks of permissible values for that node
by (pseudo-) random picks of permissible values for that node.
Works only in self.n_rows==1
"""
if not self._n_rows == 1:
raise ValueError("Replacing invalid gene values only implemented for n_rows = 1")

for gene_idx, gene_value in enumerate(dna_for_reorder):
region_idx = gene_idx // self._length_per_region
region_idx = self._get_region_idx(gene_idx)
if self._is_hidden_input_gene(gene_idx, region_idx) and gene_value > region_idx:
# replace value that is too large by smaller value
# replace value that is too large by valid value
permissible_values = self.determine_permissible_values_per_gene(gene_idx)
gene_value = rng.choice(permissible_values)
dna_for_reorder[gene_idx] = gene_value

def _determine_dependencies(self) -> Dict:
def _get_addable_nodes(self, node_dependencies: Dict) -> set:

return set(
idx for idx, dependencies in node_dependencies.items() if len(dependencies) == 0
)

def _get_region_idx(self, gene_idx):

return gene_idx // self._length_per_region

def _determine_node_dependencies(self) -> Dict:
""" Determines for every node a set of dependencies.
The set of dependencies for a node is given by the indices of the nodes,
Expand Down Expand Up @@ -416,31 +422,28 @@ def _determine_dependencies(self) -> Dict:

operator_idx = region_idx * self._length_per_region

# only consider genes which are used in current node
current_arity = self._determine_operator_arity(operator_idx)

# append each input to the dependencies
for idx_gene in range(
1, current_arity + 1
): # shift by 1 since first gene is the operator gene
input_node_idx = self._dna[operator_idx + idx_gene]
if not self._is_input_region(
input_node_idx
): # not necessary to add input regions, since they are first anyway
current_node_dependencies.add(input_node_idx) # append the input node &
current_node_dependencies.add(input_node_idx)

dependencies[str(region_idx)] = current_node_dependencies
dependencies[region_idx] = current_node_dependencies

else:
assert False # should never be reached

return dependencies

def _determine_operator_arity(self, operator_idx):
def _determine_operator_arity(self, operator_idx: int) -> int:

operator_value = self._dna[operator_idx]
arity = self._primitives[operator_value]._arity
return arity
return self._primitives[operator_value]._arity

def _permissible_inputs(self, region_idx: int) -> List[int]:

Expand Down Expand Up @@ -630,7 +633,7 @@ def mutate(self, mutation_rate: float, rng: np.random.RandomState):

for (gene_idx, allele) in zip(selected_gene_indices, np.array(dna)[selected_gene_indices]):

region_idx = gene_idx // self._length_per_region
region_idx = self._get_region_idx(gene_idx)

permissible_values = self._permissible_values[gene_idx]
permissible_alternative_values = permissible_values[permissible_values != allele]
Expand Down
8 changes: 3 additions & 5 deletions cgp/hl_api.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,10 +47,9 @@ def evolve(
not implemented. Defaults to 1.
reorder_genome: bool, optional
Whether genome reordering should be applied.
If True, reorder is applied to the parents (genome) before creating offspring.
Since the parents genome changes to a reordered genome at every generation,
this creates a neutral drift through the search space.
Defaults to False
If True, reorder is applied to the parents genomes
at every generation before creating offspring.
Reorder randomizes the genotype of an individual without changing the phenotype.
Returns
-------
None
Expand All @@ -65,7 +64,6 @@ def evolve(
# Main loop: -1 offset since the last loop iteration will still increase generation by one
while pop.generation < max_generations - 1:

# reorder the genome of the parents before creating offspring
if reorder_genome:
pop.reorder_genome()

Expand Down
3 changes: 3 additions & 0 deletions cgp/population.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -132,6 +132,9 @@ def fitness_parents(self) -> List[Union[None, float]]:
def reorder_genome(self) -> None:
""" Reorders the genome for all parents in the population
Creates a neutral drift though the search space,
by reordering parents and not offspring.
Returns
---------
None
Expand Down
73 changes: 7 additions & 66 deletions test/test_genome.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -555,66 +555,6 @@ def test_genome_reordering_empirically(rng):

pytest.importorskip("sympy")

# target: f(x_0) = x_0 ** 2 - x_0 + 1

genome_params = {
"n_inputs": 1,
"n_outputs": 1,
"n_columns": 10,
"n_rows": 1,
"levels_back": None,
"primitives": (cgp.Mul, cgp.Sub, cgp.Add, cgp.ConstantFloat),
}

genome = cgp.Genome(**genome_params)

dna_fixed = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
0, # Mul -> outs x² (address 1)
0, # x
0, # x
1, # Sub -> x² - x (address 2)
1, # x²
0, # x
1, # Sub -> outs 0 (address 3)
0, # x
0, # x
3, # const -> outs 1 (address 4)
1, # address of x² (unused)
2, # address of 0 (unused)
3, # const -> outs 1 (address 5)
0,
0,
2, # Add -> x² - x + 1 (address 6)
2, # x² - x
4, # 1
3, # const (address 7)
0,
0,
3, # const (address 8)
0,
0,
3, # const (address 9)
0,
0,
3, # const (address 10)
0,
0,
ID_OUTPUT_NODE,
6,
ID_NON_CODING_GENE,
]

genome.dna = dna_fixed
sympy_expression = cgp.CartesianGraph(genome).to_sympy()
n_reorderings = 100
for _ in range(n_reorderings):
genome.reorder(rng)
sympy_expression_after_reorder = cgp.CartesianGraph(genome).to_sympy()
assert sympy_expression_after_reorder == sympy_expression

# target: f(x_0,x_1) = x_0 ** 2 - x_1 + 1

genome_params = {
Expand All @@ -635,23 +575,23 @@ def test_genome_reordering_empirically(rng):
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
0, # Mul -> outs x² (address 2)
0, # Mul -> outs x_0² (address 2)
0, # x
0, # x
1, # Sub -> x² - y (address 3)
1, # Sub -> x_0² - x_1 (address 3)
2, # x²
1, # x
1, # y
1, # Sub -> outs 0 (address 4)
0, # x
0, # x
3, # const -> outs 1 (address 5)
2, # address of x² (unused)
2, # address of x_0² (unused)
3, # address of 0 (unused)
3, # const -> outs 1 (address 6)
0,
0,
2, # Add -> x² - x + 1 (address 7)
3, # x² - x
2, # Add -> x_0² - x_1 + 1 (address 7)
3, # x_0² - x_1
5, # 1
3, # const (address 8)
0,
Expand All @@ -672,6 +612,7 @@ def test_genome_reordering_empirically(rng):

genome.dna = dna_fixed
sympy_expression = cgp.CartesianGraph(genome).to_sympy()
n_reorderings = 100
for _ in range(n_reorderings):
genome.reorder(rng)
sympy_expression_after_reorder = cgp.CartesianGraph(genome).to_sympy()
Expand Down

0 comments on commit 50f887c

Please sign in to comment.