def apply(self):
# self.log.debug("Applying {}".format(self))
trip = self.trips[self.trip]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
trip_1 = trip[trip[:, utils.LON] < self.longitude_to_split]
trip_1 = trip_1[trip_1[:, utils.LAT].argsort()[::-1]]
trip_2 = trip[trip[:, utils.LON] >= self.longitude_to_split]
trip_2 = trip_2[trip_2[:, utils.LAT].argsort()[::-1]]
self.first_trip_percentage = len(trip_1) / len(trip)
existing_trips = [t[0, utils.TRIP] for t in self.trips]
new_trip_id = np.max(existing_trips) + 1
trip_2[:, utils.TRIP] = new_trip_id
self.trips[self.trip] = trip_1
self.trips.append(trip_2)
if self.VERIFY_COST_DELTA:
new = utils.weighted_trip_length(self.trips[self.trip][:, utils.LOCATION], self.trips[self.trip][:, utils.WEIGHT]) + \
utils.weighted_trip_length(trip_2[:, utils.LOCATION], trip_2[:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
def apply(self):
if self.trip_to_merge is None:
Neighbor.log.warning("Not applying trip merge because no valid merge was found")
return
# self.log.debug("Applying {}".format(self))
trip = self.trips[self.trip_to_merge]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(trip[:, utils.LOCATION], trip[:, utils.WEIGHT])
for trip_index in self.trip_assignments_for_gifts.keys():
old += utils.weighted_trip_length(self.trips[trip_index][:, utils.LOCATION], self.trips[trip_index][:, utils.WEIGHT])
for gift, trip_index, index_in_trip in self.gift_insertions:
gift[utils.TRIP] = self.trips[trip_index][0,1]
self.trips[trip_index] = np.insert(self.trips[trip_index], index_in_trip, gift, axis=0)
if self.VERIFY_COST_DELTA:
new = 0
for trip_index in self.trip_assignments_for_gifts.keys():
new += utils.weighted_trip_length(self.trips[trip_index][:, utils.LOCATION], self.trips[trip_index][:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new-old)
# only delete the row afterwards to not mess up the indexes for the cost calculation
del self.trips[self.trip_to_merge]
def apply(self):
# self.log.debug("Applying {}".format(self))
source = self.trips[self.trip]
destination = self.trips[self.destination_trip]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(source[:, utils.LOCATION], source[:, utils.WEIGHT]) + \
utils.weighted_trip_length(destination[:, utils.LOCATION], destination[:, utils.WEIGHT])
gift = source[
self.
gift_to_move] # NOTE: This apparently can be index-out-of-bounds!
gift[utils.TRIP] = destination[0, utils.TRIP]
destination = np.insert(destination,
self.destination_insertion_index,
gift,
axis=0)
self.trips[self.destination_trip] = destination
source = np.delete(source, self.gift_to_move, axis=0)
self.trips[self.trip] = source
if self.VERIFY_COST_DELTA:
new = utils.weighted_trip_length(source[:, utils.LOCATION], source[:, utils.WEIGHT]) + \
utils.weighted_trip_length(destination[:, utils.LOCATION], destination[:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
def apply(self):
if self.trip_to_merge is None:
Neighbor.log.warning(
"Not applying trip merge because no valid merge was found")
return
# self.log.debug("Applying {}".format(self))
trip = self.trips[self.trip_to_merge]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
for trip_index in self.modified_trips:
old += utils.weighted_trip_length(
self.trips[trip_index][:, utils.LOCATION],
self.trips[trip_index][:, utils.WEIGHT])
self.trips.clear()
self.trips.extend(self.trips_with_applied_merge)
if self.VERIFY_COST_DELTA:
new = 0
for trip_index in self.modified_trips:
new += utils.weighted_trip_length(
self.trips[trip_index][:, utils.LOCATION],
self.trips[trip_index][:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
# only delete the row afterwards to not mess up the indexes for the cost calculation
del self.trips[self.trip_to_merge]
def apply(self):
# self.log.debug("Applying {}".format(self))
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(self.trip[:, utils.LOCATION],
self.trip[:, utils.WEIGHT])
self.trip[[self.first_gift, self.second_gift
]] = self.trip[[self.second_gift, self.first_gift]]
if self.VERIFY_COST_DELTA:
new = utils.weighted_trip_length(self.trip[:, utils.LOCATION],
self.trip[:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
def print_stats(file_name=None, df=None, plots=False):
if file_name is not None:
df = pd.read_csv(file_name).merge(gifts, on="GiftId")
if df is None:
print("Need to specify either file name or df")
score = utils.weighted_reindeer_weariness(df)
trip_sizes = df.groupby("TripId").size()
trips = df.TripId.unique()
weights = np.array([df[df.TripId == trip].Weight.sum() for trip in trips])
costs = np.array([utils.weighted_trip_length(df[df.TripId == trip][["Longitude", "Latitude"]], df[df.TripId == trip].Weight) for trip in trips])
efficiencies = weights / costs
print("Score: {:.5f}B for {} trips".format(score / 1e9, len(trip_sizes)))
print("Trip sizes: min/median/max:\t\t{:>6.2f}\t{:>6.2f}\t{:>7.2f};\t{:>6.2f}+-{:>9.2f}".format(
trip_sizes.min(), trip_sizes.median(), trip_sizes.max(), trip_sizes.mean(), trip_sizes.std()**2))
print("Costs per trip: min/median/max [M]:\t{:>6.2f}\t{:>6.2f}\t{:>7.2f};\t{:>6.2f}+-{:>9.2f}".format(
costs.min()/1e6, np.median(costs)/1e6, costs.max()/1e6, costs.mean()/1e6, (costs.std()/1e6)**2))
print("Weights per trip: min/median/max:\t{:>6.2f}\t{:>6.2f}\t{:>7.2f};\t{:>6.2f}+-{:>9.2f}".format(
weights.min(), np.median(weights), weights.max(), weights.mean(), (weights.std())**2))
print("Efficiencies per trip: min/median/max:\t{:>6.2f}\t{:>6.2f}\t{:>7.2f};\t{:>6.2f}+-{:>9.2f}".format(
efficiencies.min()*1e6, np.median(efficiencies)*1e6, efficiencies.max()*1e6, efficiencies.mean()*1e6, (efficiencies.std()*1e6)**2))
if plots:
fig, axes = plt.subplots(2, 2)
axes[0, 0].hist(weights, bins=100)
axes[0, 0].set_title("Weights")
axes[0, 1].hist(costs, bins=100)
axes[0, 1].set_title("Costs")
axes[1, 0].hist(efficiencies, bins=100)
axes[1, 0].set_title("Efficiencies")
axes[1, 1].hist(trip_sizes, bins=100)
axes[1, 1].set_title("Trip sizes")
if file_name is not None:
fig.suptitle("Stats for {}".format(file_name))
def _find_best_split_index(self, trip):
minimum_cost = np.finfo(np.float64).max
best_index = None
# don't split before first item
for i in range(1, len(trip)):
first_trip = trip[:i]
second_trip = trip[i:]
cost_first_trip = utils.weighted_trip_length(
first_trip[:, utils.LOCATION], first_trip[:, utils.WEIGHT])
cost_second_trip = utils.weighted_trip_length(
second_trip[:, utils.LOCATION], second_trip[:, utils.WEIGHT])
current_cost = cost_first_trip + cost_second_trip
if current_cost < minimum_cost:
minimum_cost = current_cost
best_index = i
return best_index, minimum_cost
def apply(self):
# self.log.debug("Applying {}".format(self))
trip = self.trips[self.trip]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
new_trip = trip[self.index_to_split:]
existing_trips = [t[0, utils.TRIP] for t in self.trips]
new_trip_id = np.max(existing_trips) + 1
new_trip[:, utils.TRIP] = new_trip_id
self.trips[self.trip] = trip[:self.index_to_split]
self.trips.append(new_trip)
if self.VERIFY_COST_DELTA:
new = utils.weighted_trip_length(self.trips[self.trip][:, utils.LOCATION], self.trips[self.trip][:, utils.WEIGHT]) + \
utils.weighted_trip_length(new_trip[:, utils.LOCATION], new_trip[:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
def apply(self):
# self.log.debug("Applying {}".format(self))
first_trip = self.trips[self.first_trip]
second_trip = self.trips[self.second_trip]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(first_trip[:, utils.LOCATION], first_trip[:, utils.WEIGHT]) + \
utils.weighted_trip_length(second_trip[:, utils.LOCATION], second_trip[:, utils.WEIGHT])
# extract insertees now (before they're removed) and update their trip assignment
first_gift_row = first_trip[self.first_gift]
trip_id_for_second_gift = first_trip[0, utils.TRIP]
first_gift_row[utils.TRIP] = second_trip[0, utils.TRIP]
second_gift_row = second_trip[self.second_gift]
second_gift_row[utils.TRIP] = trip_id_for_second_gift
# update first trip
first_trip = np.insert(first_trip,
self.first_trip_insertion_index,
second_gift_row,
axis=0)
index_to_remove = self.first_gift if self.first_gift < self.first_trip_insertion_index else self.first_gift + 1
first_trip = np.delete(first_trip, index_to_remove, axis=0)
self.trips[self.first_trip] = first_trip
# update second trip
second_trip = np.insert(second_trip,
self.second_trip_insertion_index,
first_gift_row,
axis=0)
index_to_remove = self.second_gift if self.second_gift < self.second_trip_insertion_index else self.second_gift + 1
second_trip = np.delete(second_trip, index_to_remove, axis=0)
self.trips[self.second_trip] = second_trip
if self.VERIFY_COST_DELTA:
new = utils.weighted_trip_length(first_trip[:, utils.LOCATION], first_trip[:, utils.WEIGHT]) + \
utils.weighted_trip_length(second_trip[:, utils.LOCATION], second_trip[:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
def apply(self):
# self.log.debug("Applying {}".format(self))
trip = self.trips[self.trip]
if self.VERIFY_COST_DELTA:
old = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
gift = trip[self.gift_index]
trip = np.delete(trip, self.gift_index, axis=0)
index_to_insert = self.new_index if self.new_index < self.gift_index else self.new_index + 0
trip = np.insert(trip, index_to_insert, gift, axis=0)
self.trips[self.trip] = trip
if self.VERIFY_COST_DELTA:
new = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
utils.verify_costs_are_equal(self.cost_delta(), new - old)
def cost_delta(self):
if self.cost is not None:
return self.cost
trip = self.trips[self.trip]
cost_of_old_trip = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
# find split index with minimum cost
self.index_to_split, cost_of_split = self._find_best_split_index(trip)
self.cost = cost_of_split - cost_of_old_trip
return self.cost
def cost_delta(self):
if self.cost is not None:
return self.cost
trip = self.trips[self.trip]
cost_of_old_trip = utils.weighted_trip_length(trip[:, utils.LOCATION],
trip[:, utils.WEIGHT])
# check splitting in the middle third of longitudes
longitudes = np.sort(
trip[:,
utils.LON][:])[int(len(trip) / 3.0):int(len(trip) * 2.0 / 3)]
minimum_cost = np.finfo(np.float64).max
for i, lon in enumerate(longitudes):
# split trips and sort by LAT descending
trip_1 = trip[trip[:, utils.LON] < lon]
trip_1 = trip_1[trip_1[:, utils.LAT].argsort()[::-1]]
trip_2 = trip[trip[:, utils.LON] >= lon]
trip_2 = trip_2[trip_2[:, utils.LAT].argsort()[::-1]]
if len(trip_1) * len(trip_2) == 0:
# don't split here if one of the resulting trips is empty
continue
cost_2_1 = utils.weighted_trip_length(trip_1[:, utils.LOCATION],
trip_1[:, utils.WEIGHT])
cost_2_2 = utils.weighted_trip_length(trip_2[:, utils.LOCATION],
trip_2[:, utils.WEIGHT])
if cost_2_1 + cost_2_2 < minimum_cost:
minimum_cost = cost_2_1 + cost_2_2
self.longitude_to_split = lon
self.cost = minimum_cost - cost_of_old_trip
return self.cost
def evaluate_trips(self):
unique_trips = self.trips.TripId.unique()
merged = self.trips.merge(self.gifts, on="GiftId")
trips = [merged[merged.TripId == t] for t in unique_trips]
score = utils.weighted_reindeer_weariness(merged)
utils.log_success_or_error(
self.log, score < self.current_score,
"Cost of the {} trips: {:.5f}B ({:.5f}M with {} trips)".format(
unique_trips.shape[0], score / 1e9,
(score - self.current_score) / 1e6, self.current_trip_count))
utils.log_success_or_error(
self.log, score < self.current_best,
"Compared to best: {:.5f}M".format(
(score - self.current_best) / 1e6))
weights = np.asarray([trip.Weight.sum() for trip in trips])
self.log.info(
"Sleigh utilization: min {:.2f}, max {:.2f}, avg {:.2f}, std {:.2f}"
.format(weights.min(), weights.max(), weights.mean(),
weights.std()))
costs = np.asarray([
utils.weighted_trip_length(trip[["Latitude", "Longitude"]],
trip.Weight) for trip in trips
])
self.log.info(
"Trip costs: min {:.2f}M, max {:.2f}M, avg {:.2f}M, std {:.2f}k".
format(costs.min() / 1e6,
costs.max() / 1e6,
costs.mean() / 1e6,
costs.std() / 1e3))
stops = np.asarray([trip.shape[0] for trip in trips])
self.log.info(
"Stops per trip: min {}, max {}, avg {:.2f}, std {:.2f}".format(
stops.min(), stops.max(), stops.mean(), stops.std()))
cache_info = utils.get_cache_info()
self.log.info("Distance cache info: {} ({:.2f}% hits))".format(
cache_info,
100.0 * cache_info.hits / (cache_info.hits + cache_info.misses)))
def cost_delta(self):
if self.cost is not None:
return self.cost
self.trip_to_merge, self.trip_assignments_for_gifts = self._find_trip_to_merge()
if self.trip_to_merge is None:
return 0
trip = self.trips[self.trip_to_merge]
self.gift_insertions = []
cost_of_insertions = 0
for trip_index, gift in self.trip_assignments_for_gifts.items():
index_in_trip, cost = Neighbor.find_best_insertion_index(self.trips[trip_index], gift)
self.gift_insertions.append((gift, trip_index, index_in_trip))
cost_of_insertions += cost
self.cost = cost_of_insertions - utils.weighted_trip_length(trip[:, utils.LOCATION], trip[:, utils.WEIGHT])
return self.cost