354 lines
14 KiB
Python
354 lines
14 KiB
Python
from typing import List, Tuple, Dict, Union
|
|
from itertools import count
|
|
|
|
import jax
|
|
import numpy as np
|
|
from numpy.typing import NDArray
|
|
from .genome import distance
|
|
|
|
|
|
class Species(object):
|
|
|
|
def __init__(self, key, generation):
|
|
self.key = key
|
|
self.created = generation
|
|
self.last_improved = generation
|
|
self.representative: Tuple[NDArray, NDArray] = (None, None) # (nodes, connections)
|
|
self.members: List[int] = [] # idx in pop_nodes, pop_connections,
|
|
self.fitness = None
|
|
self.member_fitnesses = None
|
|
self.adjusted_fitness = None
|
|
self.fitness_history: List[float] = []
|
|
|
|
def update(self, representative, members):
|
|
self.representative = representative
|
|
self.members = members
|
|
|
|
def get_fitnesses(self, fitnesses):
|
|
return [fitnesses[m] for m in self.members]
|
|
|
|
|
|
class SpeciesController:
|
|
"""
|
|
A class to control the species
|
|
"""
|
|
|
|
def __init__(self, config):
|
|
self.config = config
|
|
self.compatibility_threshold = self.config.neat.species.compatibility_threshold
|
|
self.species_elitism = self.config.neat.species.species_elitism
|
|
self.pop_size = self.config.neat.population.pop_size
|
|
self.max_stagnation = self.config.neat.species.max_stagnation
|
|
self.min_species_size = self.config.neat.species.min_species_size
|
|
self.genome_elitism = self.config.neat.species.genome_elitism
|
|
self.survival_threshold = self.config.neat.species.survival_threshold
|
|
|
|
self.species_idxer = count(0)
|
|
self.species: Dict[int, Species] = {} # species_id -> species
|
|
|
|
self.o2m_distance_func = jax.vmap(distance, in_axes=(None, None, 0, 0)) # one to many
|
|
# self.o2o_distance_func = np_distance # one to one
|
|
self.o2o_distance_func = distance
|
|
|
|
def speciate(self, pop_nodes: NDArray, pop_connections: NDArray, generation: int) -> None:
|
|
"""
|
|
:param pop_nodes:
|
|
:param pop_connections:
|
|
:param generation: use to flag the created time of new species
|
|
:return:
|
|
"""
|
|
unspeciated = np.full((pop_nodes.shape[0],), True, dtype=bool)
|
|
previous_species_list = list(self.species.keys())
|
|
|
|
# Find the best representatives for each existing species.
|
|
new_representatives = {}
|
|
new_members = {}
|
|
|
|
for sid, species in self.species.items():
|
|
# calculate the distance between the representative and the population
|
|
r_nodes, r_connections = species.representative
|
|
distances = self.o2m_distance_wrapper(r_nodes, r_connections, pop_nodes, pop_connections)
|
|
distances = jax.device_get(distances) # fetch the data from gpu
|
|
min_idx = find_min_with_mask(distances, unspeciated) # find the min un-specified distance
|
|
|
|
new_representatives[sid] = min_idx
|
|
new_members[sid] = [min_idx]
|
|
unspeciated[min_idx] = False
|
|
|
|
# Partition population into species based on genetic similarity.
|
|
|
|
# First, fast match the population to previous species
|
|
if previous_species_list: # exist previous species
|
|
rid_list = [new_representatives[sid] for sid in previous_species_list]
|
|
res_pop_distance = [
|
|
jax.device_get(
|
|
self.o2m_distance_wrapper(pop_nodes[rid], pop_connections[rid], pop_nodes, pop_connections)
|
|
)
|
|
for rid in rid_list
|
|
]
|
|
|
|
pop_res_distance = np.stack(res_pop_distance, axis=0).T
|
|
for i in range(pop_res_distance.shape[0]):
|
|
if not unspeciated[i]:
|
|
continue
|
|
min_idx = np.argmin(pop_res_distance[i])
|
|
min_val = pop_res_distance[i, min_idx]
|
|
if min_val <= self.compatibility_threshold:
|
|
species_id = previous_species_list[min_idx]
|
|
new_members[species_id].append(i)
|
|
unspeciated[i] = False
|
|
|
|
# Second, slowly match the lonely population to new-created species.
|
|
# lonely genome is proved to be not compatible with any previous species, so they only need to be compared with
|
|
# the new representatives.
|
|
for i in range(pop_nodes.shape[0]):
|
|
if not unspeciated[i]:
|
|
continue
|
|
unspeciated[i] = False
|
|
if len(new_representatives) != 0:
|
|
# the representatives of new species
|
|
sid, rid = list(zip(*[(k, v) for k, v in new_representatives.items()]))
|
|
|
|
distances = [
|
|
self.o2o_distance_wrapper(pop_nodes[i], pop_connections[i], pop_nodes[r], pop_connections[r])
|
|
for r in rid
|
|
]
|
|
distances = np.array(distances)
|
|
min_idx = np.argmin(distances)
|
|
min_val = distances[min_idx]
|
|
if min_val <= self.compatibility_threshold:
|
|
species_id = sid[min_idx]
|
|
new_members[species_id].append(i)
|
|
continue
|
|
# create a new species
|
|
species_id = next(self.species_idxer)
|
|
new_representatives[species_id] = i
|
|
new_members[species_id] = [i]
|
|
|
|
assert np.all(~unspeciated)
|
|
|
|
# Update species collection based on new speciation.
|
|
for sid, rid in new_representatives.items():
|
|
s = self.species.get(sid)
|
|
if s is None:
|
|
s = Species(sid, generation)
|
|
self.species[sid] = s
|
|
|
|
members = new_members[sid]
|
|
s.update((pop_nodes[rid], pop_connections[rid]), members)
|
|
|
|
def update_species_fitnesses(self, fitnesses):
|
|
"""
|
|
update the fitness of each species
|
|
:param fitnesses:
|
|
:return:
|
|
"""
|
|
for sid, s in self.species.items():
|
|
# TODO: here use mean to measure the fitness of a species, but it may be other functions
|
|
s.member_fitnesses = s.get_fitnesses(fitnesses)
|
|
s.fitness = np.mean(s.member_fitnesses)
|
|
s.fitness_history.append(s.fitness)
|
|
s.adjusted_fitness = None
|
|
|
|
def stagnation(self, generation):
|
|
"""
|
|
code modified from neat-python!
|
|
:param generation:
|
|
:return: whether the species is stagnated
|
|
"""
|
|
species_data = []
|
|
for sid, s in self.species.items():
|
|
if s.fitness_history:
|
|
prev_fitness = max(s.fitness_history)
|
|
else:
|
|
prev_fitness = float('-inf')
|
|
|
|
if prev_fitness is None or s.fitness > prev_fitness:
|
|
s.last_improved = generation
|
|
|
|
species_data.append((sid, s))
|
|
|
|
# Sort in descending fitness order.
|
|
species_data.sort(key=lambda x: x[1].fitness, reverse=True)
|
|
|
|
result = []
|
|
for idx, (sid, s) in enumerate(species_data):
|
|
|
|
if idx < self.species_elitism: # elitism species never stagnate!
|
|
is_stagnant = False
|
|
else:
|
|
stagnant_time = generation - s.last_improved
|
|
is_stagnant = stagnant_time > self.max_stagnation
|
|
|
|
result.append((sid, s, is_stagnant))
|
|
return result
|
|
|
|
def reproduce(self, generation: int) -> List[Union[int, Tuple[int, int]]]:
|
|
"""
|
|
code modified from neat-python!
|
|
:param generation:
|
|
:return: crossover_pair for next generation.
|
|
# int -> idx in the pop_nodes, pop_connections of elitism
|
|
# (int, int) -> the father and mother idx to be crossover
|
|
"""
|
|
# Filter out stagnated species, collect the set of non-stagnated
|
|
# species members, and compute their average adjusted fitness.
|
|
# The average adjusted fitness scheme (normalized to the interval
|
|
# [0, 1]) allows the use of negative fitness values without
|
|
# interfering with the shared fitness scheme.
|
|
all_fitnesses = []
|
|
remaining_species = []
|
|
for stag_sid, stag_s, stagnant in self.stagnation(generation):
|
|
if not stagnant:
|
|
all_fitnesses.extend(stag_s.member_fitnesses)
|
|
remaining_species.append(stag_s)
|
|
|
|
# No species left.
|
|
if not remaining_species:
|
|
self.species = {}
|
|
return []
|
|
|
|
# Compute each species' member size in the next generation.
|
|
min_fitness = min(all_fitnesses)
|
|
max_fitness = max(all_fitnesses)
|
|
# Do not allow the fitness range to be zero, as we divide by it below.
|
|
# TODO: The ``1.0`` below is rather arbitrary, and should be configurable.
|
|
fitness_range = max(1.0, max_fitness - min_fitness)
|
|
for afs in remaining_species:
|
|
# Compute adjusted fitness.
|
|
msf = afs.fitness
|
|
af = (msf - min_fitness) / fitness_range # make adjusted fitness in [0, 1]
|
|
afs.adjusted_fitness = af
|
|
adjusted_fitnesses = [s.adjusted_fitness for s in remaining_species]
|
|
previous_sizes = [len(s.members) for s in remaining_species]
|
|
min_species_size = max(self.min_species_size, self.genome_elitism)
|
|
spawn_amounts = compute_spawn(adjusted_fitnesses, previous_sizes, self.pop_size, min_species_size)
|
|
assert sum(spawn_amounts) == self.pop_size
|
|
|
|
# generate new population and speciate
|
|
self.species = {}
|
|
# int -> idx in the pop_nodes, pop_connections of elitism
|
|
# (int, int) -> the father and mother idx to be crossover
|
|
crossover_pair: List[Union[int, Tuple[int, int]]] = []
|
|
for spawn, s in zip(spawn_amounts, remaining_species):
|
|
assert spawn >= self.genome_elitism
|
|
|
|
# retain remain species to next generation
|
|
old_members, fitnesses = s.members, s.member_fitnesses
|
|
s.members = []
|
|
self.species[s.key] = s
|
|
|
|
# add elitism genomes to next generation
|
|
sorted_members, sorted_fitnesses = sort_element_with_fitnesses(old_members, fitnesses)
|
|
if self.genome_elitism > 0:
|
|
for m in sorted_members[:self.genome_elitism]:
|
|
crossover_pair.append(m)
|
|
spawn -= 1
|
|
|
|
if spawn <= 0:
|
|
continue
|
|
|
|
# add genome to be crossover to next generation
|
|
repro_cutoff = int(np.ceil(self.survival_threshold * len(sorted_members)))
|
|
repro_cutoff = max(repro_cutoff, 2)
|
|
# only use good genomes to crossover
|
|
sorted_members = sorted_members[:repro_cutoff]
|
|
|
|
# Randomly choose parents and produce the number of offspring allotted to the species.
|
|
for _ in range(spawn):
|
|
# allow to replace, for the case that the species only has one genome
|
|
c1, c2 = np.random.choice(len(sorted_members), size=2, replace=True)
|
|
idx1, fitness1 = sorted_members[c1], sorted_fitnesses[c1]
|
|
idx2, fitness2 = sorted_members[c2], sorted_fitnesses[c2]
|
|
if fitness1 >= fitness2:
|
|
crossover_pair.append((idx1, idx2))
|
|
else:
|
|
crossover_pair.append((idx2, idx1))
|
|
|
|
return crossover_pair
|
|
|
|
def o2m_distance_wrapper(self, r_nodes, r_connections, pop_nodes, pop_connections):
|
|
# distances = self.o2m_distance_func(r_nodes, r_connections, pop_nodes, pop_connections)
|
|
# for d in distances:
|
|
# if np.isnan(d):
|
|
# print("fuck!!!!!!!!!!!!!!")
|
|
# print(distances)
|
|
# assert False
|
|
# return distances
|
|
distances = []
|
|
for nodes, connections in zip(pop_nodes, pop_connections):
|
|
d = self.o2o_distance_func(r_nodes, r_connections, nodes, connections)
|
|
if np.isnan(d) or d > 20:
|
|
np.save("too_large_distance_r_nodes.npy", r_nodes)
|
|
np.save("too_large_distance_r_connections.npy", r_connections)
|
|
np.save("too_large_distance_nodes", nodes)
|
|
np.save("too_large_distance_connections.npy", connections)
|
|
d = self.o2o_distance_func(r_nodes, r_connections, nodes, connections)
|
|
assert False
|
|
distances.append(d)
|
|
distances = np.stack(distances, axis=0)
|
|
# print(distances)
|
|
return distances
|
|
|
|
def o2o_distance_wrapper(self, *keys):
|
|
d = self.o2o_distance_func(*keys)
|
|
if np.isnan(d):
|
|
print("fuck!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
|
|
assert False
|
|
return d
|
|
|
|
|
|
def compute_spawn(adjusted_fitness, previous_sizes, pop_size, min_species_size):
|
|
"""
|
|
Code from neat-python, the only modification is to fix the population size for each generation.
|
|
Compute the proper number of offspring per species (proportional to fitness).
|
|
"""
|
|
af_sum = sum(adjusted_fitness)
|
|
|
|
spawn_amounts = []
|
|
for af, ps in zip(adjusted_fitness, previous_sizes):
|
|
if af_sum > 0:
|
|
s = max(min_species_size, af / af_sum * pop_size)
|
|
else:
|
|
s = min_species_size
|
|
|
|
d = (s - ps) * 0.5
|
|
c = int(round(d))
|
|
spawn = ps
|
|
if abs(c) > 0:
|
|
spawn += c
|
|
elif d > 0:
|
|
spawn += 1
|
|
elif d < 0:
|
|
spawn -= 1
|
|
|
|
spawn_amounts.append(spawn)
|
|
|
|
# Normalize the spawn amounts so that the next generation is roughly
|
|
# the population size requested by the user.
|
|
total_spawn = sum(spawn_amounts)
|
|
norm = pop_size / total_spawn
|
|
spawn_amounts = [max(min_species_size, int(round(n * norm))) for n in spawn_amounts]
|
|
|
|
# for batch parallelization, pop size must be a fixed value.
|
|
total_amounts = sum(spawn_amounts)
|
|
spawn_amounts[0] += pop_size - total_amounts
|
|
assert sum(spawn_amounts) == pop_size, "Population size is not stable."
|
|
|
|
return spawn_amounts
|
|
|
|
|
|
def find_min_with_mask(arr: NDArray, mask: NDArray) -> int:
|
|
masked_arr = np.where(mask, arr, np.inf)
|
|
min_idx = np.argmin(masked_arr)
|
|
return min_idx
|
|
|
|
|
|
def sort_element_with_fitnesses(members: List[int], fitnesses: List[float]) \
|
|
-> Tuple[List[int], List[float]]:
|
|
combined = zip(members, fitnesses)
|
|
sorted_combined = sorted(combined, key=lambda x: x[1], reverse=True)
|
|
sorted_members = [item[0] for item in sorted_combined]
|
|
sorted_fitnesses = [item[1] for item in sorted_combined]
|
|
return sorted_members, sorted_fitnesses
|