def fitnessfatigue_example(ff_args):
weeks = 12
goal = 1.1 * ff_args["initial_p"]
max_load = 150.0
min_load = 0.0
cycle_days = [u.WeekDays.monday, u.WeekDays.tuesday, u.WeekDays.wednesday, u.WeekDays.friday, u.WeekDays.saturday]
training_days = u.microcycle_days(cycle_days, weeks)
# training_days = u.filter_weeks(training_days, [3, 7, 11])
# print(ff_args)
solution = differential_evolution(
weeks,
goal,
training_days,
POP_SIZE,
max_load,
min_load,
ff_model.after_plan,
prequel_plan=[],
pp_func=u.sort_loads,
recomb_weight=0.7,
scale_factor=0.8,
pop_init_divisor=1,
**ff_args
)
solution_fitness = u.fitness(solution, goal, GOOD_ENOUGH_THRES, ff_model.after_plan, **ff_args)
perf_after_plan = ff_model.after_plan(solution, **ff_args)
u.print_ea_result(solution, solution_fitness, perf_after_plan, goal)
def perpot_example(pp_args):
weeks = 12
goal = 0.3
max_load = 1.0
cycle_days = [
u.WeekDays.monday, u.WeekDays.tuesday, u.WeekDays.wednesday,
u.WeekDays.friday, u.WeekDays.saturday
]
training_days = u.microcycle_days(cycle_days, weeks)
training_days = u.filter_weeks(training_days, [3, 7, 11])
# pp_args = params.pp_parms8
print(pp_args)
solution = differential_evolution(
weeks,
goal,
training_days,
POP_SIZE,
max_load,
pp_model.after_plan,
# u.sort_loads,
**pp_args)
solution_fitness = u.fitness(solution, goal, GOOD_ENOUGH_THRES,
pp_model.after_plan, **pp_args)
perf_after_plan = pp_model.after_plan(solution, **pp_args)
u.print_ea_result(solution, solution_fitness, perf_after_plan, goal)
'''
def fitnessfatigue_example(ff_args):
weeks = 12
goal = 1.1 * ff_args['initial_p']
max_load = 150.0
min_load = 0.0
cycle_days = [
u.WeekDays.monday, u.WeekDays.tuesday, u.WeekDays.wednesday,
u.WeekDays.friday, u.WeekDays.saturday
]
training_days = u.microcycle_days(cycle_days, weeks)
# training_days = u.filter_weeks(training_days, [3, 7, 11])
# print(ff_args)
solution = differential_evolution(weeks,
goal,
training_days,
POP_SIZE,
max_load,
min_load,
ff_model.after_plan,
prequel_plan=[],
pp_func=u.sort_loads,
recomb_weight=0.7,
scale_factor=0.8,
pop_init_divisor=1,
**ff_args)
solution_fitness = u.fitness(solution, goal, GOOD_ENOUGH_THRES,
ff_model.after_plan, **ff_args)
perf_after_plan = ff_model.after_plan(solution, **ff_args)
u.print_ea_result(solution, solution_fitness, perf_after_plan, goal)
def perpot_example(pp_args):
weeks = 12
goal = 0.3
max_load = 1.0
cycle_days = [u.WeekDays.monday, u.WeekDays.tuesday, u.WeekDays.wednesday, u.WeekDays.friday, u.WeekDays.saturday]
training_days = u.microcycle_days(cycle_days, weeks)
training_days = u.filter_weeks(training_days, [3, 7, 11])
# pp_args = params.pp_parms8
print(pp_args)
solution = differential_evolution(
weeks,
goal,
training_days,
POP_SIZE,
max_load,
pp_model.after_plan,
# u.sort_loads,
**pp_args
)
solution_fitness = u.fitness(solution, goal, GOOD_ENOUGH_THRES, pp_model.after_plan, **pp_args)
perf_after_plan = pp_model.after_plan(solution, **pp_args)
u.print_ea_result(solution, solution_fitness, perf_after_plan, goal)
"""
def differential_evolution(
weeks,
goal,
training_days,
pop_size,
max_load,
min_load,
model_perf_func,
prequel_plan=[], # prepend to solution candidates
pp_func=None, # post processing function
recomb_weight=0.7,
scale_factor=None, # None means dithering
pop_init_divisor=10,
**model_parameters
):
t = 0 # generation counter
fitness_t = [0.0] * pop_size
fitness_t_minus_1 = [0.0] * pop_size
local_optima_counter = 0
pops = [generate_population(weeks, training_days, pop_size, max_load, min_load, pop_init_divisor)]
if pp_func is not None:
pops[0] = list(map(pp_func, pops[0]))
good_enough = False
run_time = 0
start_time = int(time.time())
preq_plan = np.array(prequel_plan)
print("pop_init_divisor = {}".format(pop_init_divisor))
while t < 1000 and not good_enough and local_optima_counter < 40:
pops.append([0] * pop_size) # preallocate space for pointers
fitness_t_minus_1 = fitness_t.copy() # copy last run fitness values
for i in range(pop_size):
if scale_factor is None:
scale_fac = rand_scale_fac()
else:
scale_fac = scale_factor
a_i, b_i, c_i, d_i = np.random.choice(pop_size, 4, False)
a, b, c, d = pops[t][a_i], pops[t][b_i], pops[t][c_i], pops[t][d_i]
a_p = de_operator(a, b, c, d, max_load, min_load, recomb_weight, scale_fac)
if pp_func is not None:
a_p = pp_func(a_p)
a_p_fit, a_p_good_enough = u.fitness(
np.concatenate((preq_plan, a_p)), goal, GOOD_ENOUGH_THRES, model_perf_func, **model_parameters
)
a_fit, a_good_enough = u.fitness(
np.concatenate((preq_plan, a)), goal, GOOD_ENOUGH_THRES, model_perf_func, **model_parameters
)
if a_p_fit >= a_fit:
pops[t + 1][i] = a_p
fitness_t[i] = a_p_fit
if a_p_good_enough:
good_enough = True
pops[t + 1] = pops[t + 1][: i + 1] # remove superfluous entries
fitness_t = fitness_t[: i + 1] # remove superfluous entries
break
else:
pops[t + 1][i] = a
fitness_t[i] = a_fit
if a_good_enough:
good_enough = True
pops[t + 1] = pops[t + 1][: i + 1] # remove superfluous entries
fitness_t = fitness_t[: i + 1] # remove superfluous entries
break
print("gen {}: max fitness {}".format(t, max(fitness_t)))
t += 1
if max(fitness_t) == max(fitness_t_minus_1):
local_optima_counter += 1
else:
local_optima_counter = 0
run_time = int(time.time()) - start_time
s = sorted(
pops[-1],
key=lambda p: u.fitness(
np.concatenate((preq_plan, p)), goal, GOOD_ENOUGH_THRES, model_perf_func, **model_parameters
)[0],
)
print("run_time: {}".format(run_time))
print("local_optima_counter: {}".format(local_optima_counter))
print("good_enough: {}".format(good_enough))
return s[-1]
def differential_evolution(
weeks,
goal,
training_days,
pop_size,
max_load,
min_load,
model_perf_func,
prequel_plan=[], # prepend to solution candidates
pp_func=None, # post processing function
recomb_weight=0.7,
scale_factor=None, # None means dithering
pop_init_divisor=10,
**model_parameters):
t = 0 # generation counter
fitness_t = [0.0] * pop_size
fitness_t_minus_1 = [0.0] * pop_size
local_optima_counter = 0
pops = [
generate_population(weeks, training_days, pop_size, max_load, min_load,
pop_init_divisor)
]
if pp_func is not None:
pops[0] = list(map(pp_func, pops[0]))
good_enough = False
run_time = 0
start_time = int(time.time())
preq_plan = np.array(prequel_plan)
print('pop_init_divisor = {}'.format(pop_init_divisor))
while t < 1000 and not good_enough and local_optima_counter < 40:
pops.append([0] * pop_size) # preallocate space for pointers
fitness_t_minus_1 = fitness_t.copy() # copy last run fitness values
for i in range(pop_size):
if scale_factor is None:
scale_fac = rand_scale_fac()
else:
scale_fac = scale_factor
a_i, b_i, c_i, d_i = np.random.choice(pop_size, 4, False)
a, b, c, d = pops[t][a_i], pops[t][b_i], pops[t][c_i], pops[t][d_i]
a_p = de_operator(a, b, c, d, max_load, min_load, recomb_weight,
scale_fac)
if pp_func is not None:
a_p = pp_func(a_p)
a_p_fit, a_p_good_enough = u.fitness(
np.concatenate((preq_plan, a_p)), goal, GOOD_ENOUGH_THRES,
model_perf_func, **model_parameters)
a_fit, a_good_enough = u.fitness(np.concatenate(
(preq_plan, a)), goal, GOOD_ENOUGH_THRES, model_perf_func,
**model_parameters)
if a_p_fit >= a_fit:
pops[t + 1][i] = a_p
fitness_t[i] = a_p_fit
if a_p_good_enough:
good_enough = True
pops[t + 1] = pops[t + 1][:i +
1] # remove superfluous entries
fitness_t = fitness_t[:i + 1] # remove superfluous entries
break
else:
pops[t + 1][i] = a
fitness_t[i] = a_fit
if a_good_enough:
good_enough = True
pops[t + 1] = pops[t + 1][:i +
1] # remove superfluous entries
fitness_t = fitness_t[:i + 1] # remove superfluous entries
break
print('gen {}: max fitness {}'.format(t, max(fitness_t)))
t += 1
if max(fitness_t) == max(fitness_t_minus_1):
local_optima_counter += 1
else:
local_optima_counter = 0
run_time = int(time.time()) - start_time
s = sorted(pops[-1],
key=lambda p: u.fitness(np.concatenate(
(preq_plan, p)), goal, GOOD_ENOUGH_THRES, model_perf_func,
**model_parameters)[0])
print('run_time: {}'.format(run_time))
print('local_optima_counter: {}'.format(local_optima_counter))
print('good_enough: {}'.format(good_enough))
return s[-1]