day 16 python part 2 solve with A*, and cleanup
parent
bbc4075137
commit
043a57608d
185
day16/search.py
185
day16/search.py
|
@ -11,18 +11,27 @@ for line in open("input"):
|
|||
import sys, os; sys.path.append(os.path.join(os.path.dirname(__file__), "../lib"))
|
||||
import astar
|
||||
|
||||
def search():
|
||||
G2 = sorted(G, key=lambda x: (-x[1],x[0]))
|
||||
V = [v for v,_,_ in G2] # vertices
|
||||
B = {v: 1<<i for i, v in enumerate(V)} # bitmasks
|
||||
E = {B[v]: [B[e] for e in edges] for v,_,edges in G} # edge[b] -> b
|
||||
R = {B[v]: r for v,r,_ in G} # rewards: R[b] = reward
|
||||
def solve():
|
||||
G.sort(key=lambda x: (-x[1],x[0]))
|
||||
B = {v: 1<<i for i,(v,_,_) in enumerate(G)} # bitmasks
|
||||
E = {B[v]: [B[e] for e in edges] for v,_,edges in G} # E[b] = edges of b
|
||||
R = {B[v]: r for v,r,_ in G if r} # R[b] -> rate
|
||||
|
||||
all_closed = sum(B.values())
|
||||
AA = B['AA']
|
||||
|
||||
all_closed = sum(R.keys())
|
||||
all_open = 0
|
||||
|
||||
minutes = 30
|
||||
start = (B['AA'], minutes, all_closed)
|
||||
# TODO: memoize this
|
||||
def pressure(bits):
|
||||
pressure = 0
|
||||
for v,r in R.items():
|
||||
if bits&v:
|
||||
pressure += r
|
||||
return pressure
|
||||
|
||||
assert pressure(all_open) == 0
|
||||
max_pressure = pressure(all_closed)
|
||||
|
||||
# A* search minimizes costs
|
||||
# it can't maxmize anything
|
||||
|
@ -33,9 +42,13 @@ def search():
|
|||
# closed pipes instead of how much pressure is released from open pipes
|
||||
|
||||
# let's shink the graph by finding the shortest path between
|
||||
# every pair of rooms (floyd-warshall), and then building a graph which only has
|
||||
# paths from the starting room to rooms with a valve
|
||||
# and from any room with a valve to any other room with a valve
|
||||
# every pair of rooms (floyd-warshall), and then use that to build a
|
||||
# weighted graph which only has paths from any room to rooms with a valve.
|
||||
#
|
||||
# not only does this make our search space smaller,
|
||||
# it also helps by making it so that the cost changes on every step
|
||||
# (since opening a valve is the only thing that actually changes the pressure)
|
||||
# giving A* a much clearer signal about which paths are worth exploring
|
||||
dist = {}
|
||||
for v in E:
|
||||
dist[v,v] = 0
|
||||
|
@ -48,47 +61,47 @@ def search():
|
|||
if (u,t) in dist and (t,v) in dist:
|
||||
dist[u,v] = min(dist.get((u,v),999999), dist[u,t] + dist[t,v])
|
||||
|
||||
W = {v:[] for v in R if R[v]} # weighted edges
|
||||
for u in W:
|
||||
for v in W:
|
||||
W = {} # weighted edges
|
||||
for u in E:
|
||||
W[u] = []
|
||||
for v in R:
|
||||
if (u,v) in dist:
|
||||
W[u].append((v, dist[u,v]))
|
||||
aa = B['AA']
|
||||
if aa not in W:
|
||||
W[aa] = []
|
||||
for v in W:
|
||||
if v != aa and (aa,v) in dist:
|
||||
W[aa].append((v, dist[aa,v]))
|
||||
|
||||
print(W[aa])
|
||||
print(W[AA])
|
||||
|
||||
# our heuristic cost has to be <= the actual cost of getting to the goal
|
||||
#
|
||||
# here's a simple one:
|
||||
# we know it takes at least 1 minute to open a valve,
|
||||
# and at least another minute to walk to the valve
|
||||
# so we can assign at least cost 2*pressure(closed_valves) to this node
|
||||
# (unless there is only 1 minute left)
|
||||
#
|
||||
# we can keep doing that until there are no valves left to open
|
||||
# or there is no time left.
|
||||
#
|
||||
# note that the nodes with the largest flow rate are assigned
|
||||
# the lowest position in the bitmap, so clearing the bits from
|
||||
# low to high will always give us the optimal order
|
||||
def heuristic(node):
|
||||
v, minutes, closed = node
|
||||
m = min(2, minutes)
|
||||
return m*pressure(closed)
|
||||
# assume we can open a valve every 2 minutes
|
||||
# how much would that cost?
|
||||
c = 0
|
||||
while closed and minutes > 0:
|
||||
c += pressure(closed) * min(minutes,2)
|
||||
closed &= (closed-1)
|
||||
minutes -= 2
|
||||
return c
|
||||
|
||||
def pressure(bits):
|
||||
pressure = 0
|
||||
for v,r in R.items():
|
||||
if bits&v:
|
||||
pressure += r
|
||||
return pressure
|
||||
|
||||
assert pressure(all_open) == 0
|
||||
|
||||
def is_goal(n):
|
||||
v, minutes, closed = n
|
||||
def is_goal(node):
|
||||
v, minutes, closed = node
|
||||
return minutes == 0 or closed == all_open
|
||||
|
||||
info = {}
|
||||
def neighbors(n):
|
||||
v, minutes, closed = n
|
||||
def neighbors(node):
|
||||
v, minutes, closed = node
|
||||
if minutes not in info:
|
||||
print(info)
|
||||
info[minutes] = 0
|
||||
|
@ -103,7 +116,7 @@ def search():
|
|||
can_move = False
|
||||
for e, dist in W[v]:
|
||||
t = dist + 1
|
||||
if e&closed and t <= minutes and R[e]:
|
||||
if e&closed and t <= minutes:
|
||||
can_move = True
|
||||
c = pressure(closed)*t
|
||||
yield c, (e, minutes-t, closed&~e)
|
||||
|
@ -111,17 +124,93 @@ def search():
|
|||
if can_move == False:
|
||||
yield pressure(closed)*minutes, (v, 0, closed)
|
||||
|
||||
minutes = 30
|
||||
start = (AA, minutes, all_closed)
|
||||
|
||||
c, _, path = astar.search(start, is_goal, neighbors, heuristic)
|
||||
print(c)
|
||||
print(pressure(all_closed)*minutes - c)
|
||||
print(max_pressure*minutes - c)
|
||||
|
||||
#maxpair = []
|
||||
#def pairs():
|
||||
# O = sorted(best.keys())
|
||||
# for i in range(len(O)):
|
||||
# for j in range(i,len(O)):
|
||||
# if not O[i] & O[j]:
|
||||
# yield(best[O[i]]+best[O[j]])
|
||||
#print(max(pairs()))
|
||||
def heuristic2(node):
|
||||
if is_goal2(node):
|
||||
return 0
|
||||
v1, v2, min1, min2, closed = node
|
||||
# if the players are out of sync, assume the other player
|
||||
# will close one valve when they catch up
|
||||
if min1 != min2:
|
||||
closed &= closed - 1
|
||||
# assume we can open a valve every minute remaining
|
||||
# how much would that cost?
|
||||
c = 0
|
||||
pr = pressure(closed)
|
||||
m = min(min1, min2)
|
||||
while closed and m > 0:
|
||||
c += pr
|
||||
tmp = closed
|
||||
closed &= (closed-1)
|
||||
pr -= R[tmp-closed]
|
||||
m -= 1
|
||||
return c
|
||||
|
||||
search()
|
||||
def is_goal2(node):
|
||||
_, _, min1, min2, closed = node
|
||||
return min1 == 0 and min2 == 0 or closed == all_open
|
||||
|
||||
def neighbors2(node):
|
||||
v1, v2, min1, min2, closed = node
|
||||
|
||||
if min(min1,min2) not in info:
|
||||
print(info)
|
||||
info.setdefault(min1, 0)
|
||||
info.setdefault(min2, 0)
|
||||
info[min1] += 1
|
||||
info[min2] += 1
|
||||
|
||||
if min1 <= 0 and min2 <= 0:
|
||||
pass
|
||||
elif closed == all_open:
|
||||
yield 0, (v1, v2, 0, 0, closed)
|
||||
else:
|
||||
moved = False
|
||||
# either player can move
|
||||
# but we can't open a valve that would take less time to open
|
||||
# than the other player has already used. we've already paid
|
||||
# the cost for _not_ opening those valves and we can't
|
||||
# retroactively change that (no time travel)
|
||||
|
||||
# if both players are in the same spot and have the same amount of
|
||||
# time remaining, then only let one of them move in order to break
|
||||
# symmetries (this can only happen in the start state)
|
||||
|
||||
# TODO: are there more symmetries we can break?
|
||||
|
||||
# move to a closed valve and open it
|
||||
discount1 = max(min1-min2, 0)
|
||||
for e, dist in W[v1]:
|
||||
t = dist + 1
|
||||
if e&closed and discount1 <= t <= min1:
|
||||
moved = True
|
||||
c = pressure(closed)*(t-discount1)
|
||||
yield c, (e, v2, min1-t, min2, closed&~e)
|
||||
if (v1, min1) != (v2, min2):
|
||||
discount2 = max(min2-min1, 0)
|
||||
for e, dist in W[v2]:
|
||||
t = dist + 1
|
||||
if e&closed and discount2 <= t <= min2:
|
||||
moved = True
|
||||
c = pressure(closed)*(t-discount2)
|
||||
yield c, (v1, e, min1, min2-t, closed&~e)
|
||||
|
||||
# are there no moves left?
|
||||
# then wait out the timer
|
||||
if not moved:
|
||||
yield pressure(closed)*min(min1,min2), (v1, v2, 0, 0, closed)
|
||||
|
||||
minutes = 26
|
||||
start2 = (AA, AA, minutes, minutes, all_closed)
|
||||
info.clear()
|
||||
c, _, path = astar.search(start2, is_goal2, neighbors2, heuristic2)
|
||||
print(c)
|
||||
print(max_pressure*minutes - c)
|
||||
|
||||
solve()
|
||||
|
|
Loading…
Reference in New Issue