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 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 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 getAction(self, game_state: GameState) -> str:
"""
Returns the minimax action from the current gameState using
self.depth and self.evaluationFunction.
Here are some method calls that might be useful when implementing
minimax.
gameState.getLegalActions(agentIndex): Returns a list of legal
actions for an agent agentIndex=0 means Pacman, ghosts are >= 1
gameState.generateSuccessor(agentIndex, action): Returns the
successor game state after an agent takes an action
gameState.getNumAgents(): Returns the total number of agents in the
game """
legal_actions = game_state.getLegalActions(agentIndex=0)
best_action_index = max(range(len(legal_actions)),
key=lambda action_num: self.min_value(
state=game_state.generateSuccessor(
agentIndex=0,
action=legal_actions[action_num],
),
depth=self.depth,
ghost_num=1,
))
return legal_actions[best_action_index]
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 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 _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 getAction(self, gameState: GameState) -> str:
"""
Returns the minimax action using self.depth and self.evaluationFunction
"""
# BEGIN_YOUR_CODE (our solution is 36 lines of code, but don't worry if you deviate from this)
def getVal(s,
d,
agentIndex,
alpha=float('-inf'),
beta=float('inf'),
evalFn=self.evaluationFunction):
nextAgentIndex = 0 if agentIndex == s.getNumAgents(
) - 1 else agentIndex + 1
actions = s.getLegalActions(agentIndex)
if len(actions) == 0:
return s.getScore()
elif d == 0:
if agentIndex != 0:
raise Exception(
f"Unexpected agentIndex {agentIndex} != {0}")
return evalFn(s)
elif agentIndex == 0:
maxVal = float('-inf')
# actions.sort(key=lambda a: evalFn(s.generateSuccessor(agentIndex, a)), reverse=True)
for a in actions:
maxVal = max(
maxVal,
getVal(s.generateSuccessor(agentIndex, a), d,
nextAgentIndex, alpha, beta))
alpha = max(alpha, maxVal)
if alpha >= beta:
break
return maxVal
else:
nextD = d - (1 if agentIndex == s.getNumAgents() - 1 else 0)
minVal = float('inf')
# actions.sort(key=lambda a: evalFn(s.generateSuccessor(agentIndex, a)), reverse=False)
for a in actions:
minVal = min(
minVal,
getVal(s.generateSuccessor(agentIndex, a), nextD,
nextAgentIndex, alpha, beta))
beta = min(beta, minVal)
if alpha >= beta:
break
return minVal
targetVal = getVal(gameState, self.depth, 0)
# print(f"AlphaBetaAgent value of state = {targetVal}")
legalActions = gameState.getLegalActions(0)
actions = [
a for a in legalActions if getVal(
gameState.generateSuccessor(0, a), self.depth, 1) == targetVal
]
return random.choice(actions)
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 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 getAction(self, gameState: GameState):
"""
Returns the minimax action from the current gameState using self.depth
and self.evaluationFunction.
"""
_max = float("-inf")
action = None
for move in gameState.getLegalActions(0):
util = minimax(self.evaluationFunction, 1, 0,
gameState.generateSuccessor(0, move), self.depth)
if util > _max or _max == float("-inf"):
_max = util
action = move
return action
def getAction(self, game_state: GameState) -> str:
"""
Returns the minimax action from the current gameState using
self.depth and self.evaluationFunction.
Here are some method calls that might be useful when implementing
minimax.
gameState.getLegalActions(agentIndex): Returns a list of legal
actions for an agent agentIndex=0 means Pacman, ghosts are >= 1
gameState.generateSuccessor(agentIndex, action): Returns the
successor game state after an agent takes an action
gameState.getNumAgents(): Returns the total number of agents in the
game
"""
legal_actions = game_state.getLegalActions(agentIndex=0)
alpha, beta = -inf, inf
utility = -inf
for action_num in range(len(legal_actions)):
successor = game_state.generateSuccessor(
agentIndex=0,
action=legal_actions[action_num],
)
utility = max(
utility,
self.min_value(
successor,
depth=self.depth,
alpha=alpha,
beta=beta,
ghost_num=1,
),
)
if utility > alpha:
best_action_index = action_num
alpha = utility
return legal_actions[best_action_index]
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 getAction(self, gameState: GameState) -> str:
"""
Returns the expectimax action using self.depth and self.evaluationFunction
All ghosts should be modeled as choosing uniformly at random from their
legal moves.
"""
# BEGIN_YOUR_CODE (our solution is 20 lines of code, but don't worry if you deviate from this)
def getVal(s, d, agentIndex, evalFn=self.evaluationFunction):
nextAgentIndex = 0 if agentIndex == s.getNumAgents(
) - 1 else agentIndex + 1
actions = s.getLegalActions(agentIndex)
if len(actions) == 0:
return s.getScore()
elif d == 0:
if agentIndex != 0:
raise Exception(
f"Unexpected agentIndex {agentIndex} != {0}")
return evalFn(s)
elif agentIndex == 0:
return max(
getVal(s.generateSuccessor(agentIndex, a), d,
nextAgentIndex) for a in actions)
else:
nextD = d - (1 if agentIndex == s.getNumAgents() - 1 else 0)
return sum((1 / len(actions)) * getVal(
s.generateSuccessor(agentIndex, a), nextD, nextAgentIndex)
for a in actions)
targetVal = getVal(gameState, self.depth, 0)
# print(f"MinimaxAgent value of state = {targetVal}")
legalActions = gameState.getLegalActions(0)
actions = [
a for a in legalActions if getVal(
gameState.generateSuccessor(0, a), self.depth, 1) == targetVal
]
return random.choice(actions)
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 getAction(self, gameState: GameState) -> str:
"""
Returns the minimax action from the current gameState using self.depth
and self.evaluationFunction. Terminal states can be found by one of the following:
pacman won, pacman lost or there are no legal moves.
Here are some method calls that might be useful when implementing minimax.
gameState.getLegalActions(agentIndex):
Returns a list of legal actions for an agent
agentIndex=0 means Pacman, ghosts are >= 1
gameState.generateSuccessor(agentIndex, action):
Returns the successor game state after an agent takes an action
gameState.getNumAgents():
Returns the total number of agents in the game
gameState.getScore():
Returns the score corresponding to the current state of the game
gameState.isWin():
Returns True if it's a winning state
gameState.isLose():
Returns True if it's a losing state
self.depth:
The depth to which search should continue
"""
# BEGIN_YOUR_CODE (our solution is 20 lines of code, but don't worry if you deviate from this)
def getVal(s, d, agentIndex, evalFn=self.evaluationFunction):
nextAgentIndex = 0 if agentIndex == s.getNumAgents(
) - 1 else agentIndex + 1
actions = s.getLegalActions(agentIndex)
if len(actions) == 0:
return s.getScore()
elif d == 0:
if agentIndex != 0:
raise Exception(
f"Unexpected agentIndex {agentIndex} != {0}")
return evalFn(s)
elif agentIndex == 0:
return max(
getVal(s.generateSuccessor(agentIndex, a), d,
nextAgentIndex) for a in actions)
else:
nextD = d - (1 if agentIndex == s.getNumAgents() - 1 else 0)
return min(
getVal(s.generateSuccessor(agentIndex, a), nextD,
nextAgentIndex) for a in actions)
targetVal = getVal(gameState, self.depth, 0)
# print(f"MinimaxAgent value of state = {targetVal}")
legalActions = gameState.getLegalActions(0)
actions = [
a for a in legalActions if getVal(
gameState.generateSuccessor(0, a), self.depth, 1) == targetVal
]
return random.choice(actions)