def min_value(self, game_state: GameState, depth=0, ghost_index=0, alpha=-math.inf, beta=math.inf):
# next_ghost_to_move
ghost_index += 1
if self.is_a_new_level_of_search(game_state, ghost_index):
depth = depth + 1
if game_state.isWin() or game_state.isLose():
return self.evaluationFunction(game_state)
value = math.inf
legal_actions = game_state.getLegalActions(ghost_index)
for action in legal_actions:
successor = game_state.getNextState(ghost_index, action)
if self.is_a_new_level_of_search(game_state, ghost_index):
# let's move on with pacman since this is the last agent (new max node)
value = min(value, self.max_value(successor, depth=depth, alpha=alpha, beta=beta))
else:
# next on the tree is another minimizer, lets continue with another ghost
value = min(value,
self.min_value(successor, depth=depth, ghost_index=ghost_index, alpha=alpha, beta=beta))
if value < alpha:
return value
beta = min(beta, value)
return value
def exp_value(self, game_state: GameState, depth=0, ghost_index=0):
# next_ghost_to_move
ghost_index += 1
if self.is_a_new_level_of_search(game_state, ghost_index):
depth = depth + 1
if game_state.isWin() or game_state.isLose():
return self.evaluationFunction(game_state)
value = 0
legal_actions = game_state.getLegalActions(ghost_index)
for action in legal_actions:
successor = game_state.getNextState(ghost_index, action)
probability = 1 / len(legal_actions)
if self.is_a_new_level_of_search(game_state, ghost_index):
# let's move on with pacman since this is the last agent (new max node)
value += probability * self.max_value(successor, depth=depth)
else:
# next on the tree is another minimizer, lets continue with another ghost
value += probability * self.exp_value(successor, depth=depth, ghost_index=ghost_index)
return value
def min_value(self, state: GameState, depth: int, ghost_num: int) -> int:
# Game over or search depth has been reached
if state.isLose() or state.isWin() or depth <= 0:
return self.evaluationFunction(state)
# Sanity check: valid ghost number?
assert 1 <= ghost_num < state.getNumAgents()
legal_actions = state.getLegalActions(ghost_num)
successors = [
state.generateSuccessor(ghost_num, ghost_action)
for ghost_action in legal_actions
]
# If this is the last ghost, next optimizer should be from pacman's
# perspective
next_optimizer = self.max_value \
if ghost_num == state.getNumAgents() - 1 \
else self.min_value
# If this is the last ghost, decrement depth
next_depth = depth - 1 \
if ghost_num == state.getNumAgents() - 1 \
else depth
utilities = [
next_optimizer(state, next_depth, ghost_num + 1)
for state in successors
]
return min(utilities)
def searchTree(state: GameState, depth: int, agent: int):
actions = state.getLegalActions(agent)
nextAgent = (agent + 1) % state.getNumAgents()
if state.isLose() or state.isWin() or len(actions) == 0:
return [state.getScore(), None]
elif depth == 0:
return [self.evaluationFunction(state), None]
elif agent == 0:
successors = [
searchTree(state.generateSuccessor(agent, action), depth,
nextAgent)[0] for action in actions
]
maximum = max(successors)
maxIndex = successors.index(maximum)
return [maximum, actions[maxIndex]]
else:
nextDepth = depth
if nextAgent == 0:
nextDepth -= 1
successors = [
searchTree(state.generateSuccessor(agent, action),
nextDepth, nextAgent)[0] for action in actions
]
expected = sum(successors) * 1.0 / len(successors)
return [expected, None]
def max_value(
self,
state: GameState,
depth: int,
alpha: int,
beta: int,
actor: Optional[int] = None,
) -> int:
# Sanity check: have all the ghosts been evaluated the last round?
if actor is not None:
assert actor == state.getNumAgents()
# Game over or search depth has been reached
if state.isLose() or state.isWin() or depth <= 0:
return self.evaluationFunction(state)
legal_actions = state.getLegalActions(agentIndex=0)
utility = -inf
for action in legal_actions:
successor = state.generateSuccessor(agentIndex=0, action=action)
utility = max(
utility,
self.min_value(successor, depth, alpha, beta, ghost_num=1),
)
if utility > beta:
return utility
alpha = max(alpha, utility)
return utility
def _alphabeta(self, gameState: GameState, idx: int, ab: List[float]) -> Tuple[float, str]:
n = gameState.getNumAgents()
if idx / n >= self.depth or gameState.isWin() or gameState.isLose():
return (self.evaluationFunction(gameState), None)
agent = idx % n
legalActions = gameState.getLegalActions(agent)
pacman = (agent == 0)
idx0 = int(pacman)
idx1 = int(not pacman)
mod = 1 if pacman else -1
best_score = -float('inf') * mod
best_action = None
for legalAction in legalActions:
s = gameState.generateSuccessor(agent, legalAction)
score = self._alphabeta(s, idx + 1, [*ab])[0]
if score * mod > best_score * mod:
best_score, best_action = score, legalAction
if best_score * mod > ab[idx0] * mod:
break
ab[idx1] = max(ab[idx1] * mod, best_score * mod) * mod
return (best_score, best_action)
def rb_directional_expectimax(self, cur_state: GameState, turn: int,
agent: int, depth_limit: int, depth: int,
ghost_num: int):
if turn == agent:
depth += 1
if depth >= depth_limit or cur_state.isWin() or cur_state.isLose():
return self.evaluationFunction(cur_state)
if turn == agent: # if Pacman's turn
cur_max = np.NINF
for action in cur_state.getLegalPacmanActions(
): # iterating over children gameStates
child_state = cur_state.generateSuccessor(turn, action)
cur_max = max(
cur_max,
self.rb_directional_expectimax(
child_state, (turn + 1) % (ghost_num + 1), agent,
depth_limit, depth, ghost_num))
return cur_max
else: # if ghost turn
assert turn > agent
ghost_legal_moves = cur_state.getLegalActions(turn)
ghost = DirectionalGhost(turn)
# assert len(ghost_legal_moves) is not 0
expectancy = 0
for action in ghost_legal_moves:
child_state = cur_state.generateSuccessor(turn, action)
dist = ghost.getDistribution(cur_state)
# print(dist)
expectancy += (dist[action]) * (self.rb_directional_expectimax(
child_state, (turn + 1) %
(ghost_num + 1), agent, depth_limit, depth, ghost_num))
if math.isnan(expectancy):
expectancy = 0
return expectancy
def minimax(evalFunc: classmethod, agent: int, depth: int, gameState: GameState, maxDepth: int) -> float:
if gameState.isLose() or gameState.isWin() or depth == maxDepth:
return evalFunc(gameState)
if agent == 0:
return max(minimax(evalFunc, 1, depth, gameState.generateSuccessor(agent, state), maxDepth) for state in gameState.getLegalActions(agent))
else:
nextAgent = agent + 1
if gameState.getNumAgents() == nextAgent:
nextAgent = 0
if nextAgent == 0:
depth += 1
return min(minimax(evalFunc, nextAgent, depth, gameState.generateSuccessor(agent, state), maxDepth) for state in gameState.getLegalActions(agent))
def min_value(
self,
state: GameState,
depth: int,
alpha: int,
beta: int,
ghost_num: int,
) -> int:
# Game over or search depth has been reached
if state.isLose() or state.isWin() or depth <= 0:
return self.evaluationFunction(state)
# Sanity check: valid ghost number?
assert 1 <= ghost_num < state.getNumAgents()
legal_actions = state.getLegalActions(ghost_num)
# If this is the last ghost, next optimizer should be from pacman's
# perspective
next_optimizer = self.max_value \
if ghost_num == state.getNumAgents() - 1 \
else self.min_value
# If this is the last ghost, decrement depth
next_depth = depth - 1 if ghost_num == state.getNumAgents(
) - 1 else depth
utility = inf
for action in legal_actions:
successor = state.generateSuccessor(
agentIndex=ghost_num,
action=action,
)
utility = min(
utility,
next_optimizer(
successor,
next_depth,
alpha,
beta,
ghost_num + 1,
),
)
if utility < alpha:
return utility
beta = min(beta, utility)
return utility
def betterEvaluationFunction(currentGameState:GameState):
"""
Your extreme ghost-hunting, pellet-nabbing, food-gobbling, unstoppable
evaluation function (question 5).
DESCRIPTION: <write something here so we know what you did>
"""
features = [0.]*8
weights = [20000,10,1,1,45,0.5,3,1]
#1/number_food,1/cloest_food_dis,
#1/num_pallets, 1/cloest_pallet, if_empty_pallet(encourage to eat the last pallet), scared_time(encourage the pacman to eat pallet to increase scared_time)
#encourage to eat the ghost, some scared time becomes 0, some are not
#1/distance_to_ghost
#weights = []
if currentGameState.isWin():
#print(currentGameState.getFood().asList())
return float("inf")
if currentGameState.isLose():
return float("-inf")
pac_pos = currentGameState.getPacmanPosition()
food_pos = currentGameState.getFood().asList()
num_food = len(food_pos)
ghost_states = currentGameState.getGhostStates()
ghost_scared_times = [ghostState.scaredTimer for ghostState in ghost_states]
ghost_pos = [g.getPosition() for g in ghost_states]
pallets = currentGameState.getCapsules()
features[0] = 1/num_food
food_dis = sorted([util.manhattanDistance(fp,pac_pos) for fp in food_pos])
features[1] = 1/food_dis[0]
pallets_dis = sorted([util.manhattanDistance(p,pac_pos) for p in pallets])
if pallets_dis:
features[2] = 1/len(pallets_dis)
features[3] = 1/pallets_dis[0]
features[4] = int(not pallets) # if pallets become zero, we need to encourage this
features[5] = sum(ghost_scared_times)
features[6] = 1 if 0 in ghost_scared_times and sum(ghost_scared_times)>0 else 0
# if some ghost's scared time becomes zero, some are not, then some ghost is eaten
# for those states after the ghost is eaten, this will always be 1, but does not afftect the action selection, because we compare the relative relation
ghost_dis = [util.manhattanDistance(g, pac_pos) for g in ghost_pos]
minDis = min(ghost_dis)
min_dis_ghost_inds = [i for i in range(len(ghost_dis)) if ghost_dis[i] == minDis]
min_dis_ghost_scared_times = [ghost_scared_times[i] for i in min_dis_ghost_inds]
if any(st < minDis for st in min_dis_ghost_scared_times):
features[7] = -1.0 / minDis # if any closet ghost has no enough scared time, then the agent is in danger
else:
features[7] = 1.0 / minDis # otherwise, there is no danger
return sum(w*f for w,f in zip(weights,features))
def _minimax(self, gameState: GameState, idx: int) -> Tuple[float, str]:
n = gameState.getNumAgents()
if idx / n >= self.depth or gameState.isWin() or gameState.isLose():
return (self.evaluationFunction(gameState), None)
agent = idx % n
legalActions = gameState.getLegalActions(agent)
mod = 1 if agent == 0 else -1
best_score = -float('inf') * mod
best_action = None
for legalAction in legalActions:
s = gameState.generateSuccessor(agent, legalAction)
score = self._minimax(s, idx + 1)[0]
if score * mod > best_score * mod:
best_score, best_action = score, legalAction
return (best_score, best_action)
def max_value(self,
state: GameState,
depth: int,
actor: Optional[int] = None) -> int:
# Sanity check: have all the ghosts been evaluated the last round?
if actor is not None:
assert actor == state.getNumAgents()
# Game over or search depth has been reached
if state.isLose() or state.isWin() or depth <= 0:
return self.evaluationFunction(state)
legal_actions = state.getLegalActions(agentIndex=0)
successors = [
state.generateSuccessor(agentIndex=0, action=action)
for action in legal_actions
]
utilities = [
self.min_value(state, depth, ghost_num=1) for state in successors
]
return max(utilities)
def _expectimax(self, gameState: GameState, idx: int) -> Tuple[float, str]:
n = gameState.getNumAgents()
if idx / n >= self.depth or gameState.isWin() or gameState.isLose():
return (self.evaluationFunction(gameState), None)
agent = idx % n
legalActions = gameState.getLegalActions(agent)
n_actions = len(legalActions)
ret_score = -float('inf') if agent == 0 else 0
ret_action = None
for legalAction in legalActions:
s = gameState.generateSuccessor(agent, legalAction)
score = self._expectimax(s, idx + 1)[0]
if agent != 0:
ret_score += score / n_actions
elif score > ret_score:
ret_score, ret_action = score, legalAction
return (ret_score, ret_action)
def betterEvaluationFunction(currentGameState: GameState):
"""
Your extreme ghost-hunting, pellet-nabbing, food-gobbling, unstoppable
evaluation function (question 5).
DESCRIPTION: <write something here so we know what you did>
"""
"*** YOUR CODE HERE ***"
# 首先判断当前是否已经结束
if currentGameState.isWin():
return float('inf')
if currentGameState.isLose():
return -float('inf')
score = currentGameState.getScore()
# 考虑food
foods = currentGameState.getFood().asList()
foodDis = [
util.manhattanDistance(currentGameState.getPacmanPosition(), food)
for food in foods
]
foodDis.sort(reverse=True)
numFoods = 3
if currentGameState.getNumFood() < 3:
numFoods = currentGameState.getNumFood()
score -= foodDis[0] * 1.5
# for i in range(numFoods):
# score-=(numFoods-i)*foodDis[i]
# 如果附近有food 最好可以吃到
score -= (currentGameState.getNumFood() * 4)
# 考虑ghost
ghostDis = [
util.manhattanDistance(currentGameState.getPacmanPosition(), ghost)
for ghost in currentGameState.getGhostPositions()
]
score += max(3, min(ghostDis)) * 2
# 考虑capsule
score -= len(currentGameState.getCapsules()) * 4
return score
def searchTree(state: GameState, depth: int, agent: int, a, b):
actions = state.getLegalActions(agent)
nextAgent = (agent + 1) % state.getNumAgents()
if state.isLose() or state.isWin() or len(actions) == 0:
return [state.getScore(), None]
elif depth == 0:
return [self.evaluationFunction(state), None]
elif agent == 0:
value = float('-inf')
successors = []
for action in actions:
curr = searchTree(state.generateSuccessor(agent, action),
depth, nextAgent, a, b)[0]
successors.append(curr)
value = max(value, curr)
a = max(a, value)
if a >= b:
break
maxIndex = successors.index(value)
return [value, actions[maxIndex]]
else:
nextDepth = depth
if nextAgent == 0:
nextDepth -= 1
value = float('inf')
successors = []
for action in actions:
curr = searchTree(state.generateSuccessor(agent, action),
nextDepth, nextAgent, a, b)[0]
successors.append(curr)
value = min(value, curr)
b = min(b, value)
if a >= b:
break
minIndex = successors.index(value)
return [value, actions[minIndex]]
def is_terminal_state(self, game_state: GameState, current_depth):
return game_state.isWin() or game_state.isLose() or current_depth == self.depth