def measure_efficiency(evolution, n_preys, n_tests, confusion):
total_preys, total_visible_preys, total_successes, attempts = list(), list(
), list(), list()
if confusion:
for t in range(n_tests):
grid = Grid(512, 512)
predator = evolution.predator_pool[0]
Predator(grid,
brain=predator.brain,
mask=predator.mask,
confusion=True)
prey = evolution.prey_pool[0]
for _ in range(n_preys):
Prey(grid, brain=prey.brain)
print("\nRunning test", t + 1, "out of", n_tests)
visible_preys, successes = list(), list()
for _ in range(2000):
for animal in grid.inhabitants:
result = animal.look_around()
if result:
visible_preys.append(result[0])
successes.append(result[1])
for animal in grid.inhabitants:
animal.pick_move()
#print(grid.n_preys(),'preys left')
#print('Average number of visible preys in attack:',np.mean(visible_preys))
#print('Proportion of successes:',len(np.where(successes)[0])/len(successes))
total_preys.append(grid.n_preys())
total_visible_preys.append(np.mean(visible_preys))
total_successes.append(
len(np.where(successes)[0]) / len(successes))
attempts.append(len(successes))
print('\nAverage number of preys left:', np.mean(total_preys))
print('\nAverage number of visible preys in attacks:',
np.mean(total_visible_preys))
print('\nProportion of successes:', np.mean(total_successes))
print('\nAverage number of attacks:', np.mean(attempts))
else:
preys_left = list()
for t in range(n_tests):
grid = Grid(512, 512)
predator = evolution.predator_pool[0]
Predator(grid,
brain=predator.brain,
mask=predator.mask,
confusion=False)
prey = evolution.prey_pool[0]
for _ in range(n_preys):
Prey(grid, brain=prey.brain)
print("\nRunning test", t + 1, "out of", n_tests)
for _ in range(2000):
for animal in grid.inhabitants:
animal.look_around()
for animal in grid.inhabitants:
animal.pick_move()
preys_left.append(grid.n_preys())
print('\nAverage number of preys left:', np.mean(preys_left))
def __init__(self, Environment, location):
self.Environment = Environment
# Initialize Q
alpha = 0.3
gamma = 0.7
epsilon = 0.1
self.Prey = Prey(Environment, location)
self.actions = Prey.actions
self.TeamQLearning = TeamQLearning(self, alpha, gamma, epsilon)
def update_prey_pool(self, ranked_preys):
# same syntax as update_predator_pool but with preys
kids, p1, p2 = self.make_up_next_prey_generation(ranked_preys)
self.prey_pool = [
Prey(Grid(*self.grid_size), brain=p1),
Prey(Grid(*self.grid_size), brain=p2)
]
for kid in kids:
self.prey_pool.append(kid)
def move(self, directionX, directionY):
oldXPosition = self.gridX
oldYPosition = self.gridY
nextXPosition = self.gridX + directionX
nextYPosition = self.gridY + directionY
if self.isMovementPossible(nextXPosition, nextYPosition):
# If prey is in new position
if not PreyAdult.dictionaryOfPreyAdults.has_key((nextXPosition, nextYPosition)):
# Another PreyAdult is not on the next Position, so valid movement
Prey.move(self, directionX, directionY)
PreyAdult.dictionaryOfPreyAdults.pop((oldXPosition, oldYPosition))
PreyAdult.dictionaryOfPreyAdults[(self.gridX, self.gridY)] = self
def move(self, directionX, directionY):
oldXPosition = self.gridX
oldYPosition = self.gridY
nextXPosition = self.gridX + directionX
nextYPosition = self.gridY + directionY
if (self.isMovementPossible(nextXPosition, nextYPosition)):
#If prey is in new position
if (not PreyAdult.dictionaryOfPreyAdults.has_key(
(nextXPosition, nextYPosition))):
#Another PreyAdult is not on the next Position, so valid movement
Prey.move(self, directionX, directionY)
PreyAdult.dictionaryOfPreyAdults.pop(
(oldXPosition, oldYPosition))
PreyAdult.dictionaryOfPreyAdults[(self.gridX,
self.gridY)] = self
def test_one_on_one(predator, prey, return_dict, num_test, pr):
grid = Grid(512, 512)
# copy the predator
test_predator = Predator(grid,
brain=predator.brain,
confusion=self.confusion,
mask=self.mask)
# copy and duplicate the prey (n_preys instances)
swarm = [Prey(grid, brain=prey.brain) for _ in range(self.n_preys)]
for i in range(self.n_iters):
# next 4 lines to print the % of progress in the simulation
#if i%(self.n_iters/10)==0:
#if pr[int(i/(self.n_iters/10))]<self.card_prey_pool*self.card_predator_pool-10:
# pr[int(i/(self.n_iters/10))]+=1
#else:
# t = np.round((time.time()-start)/60,2)
# t = str(t).split('.')
# t = [t[0],str(int(int(t[1])*.6))]
# print(int(i*100/self.n_iters),'%\t'+t[0]+'m'+t[1]+'s')
# pr[int(i/(self.n_iters/10))] = -self.card_prey_pool*self.card_predator_pool
self.next_timeStep(grid)
return_dict[num_test] = {
'predator': test_predator.fitness,
'prey': swarm[0].fitness
}
def make_up_next_prey_generation(self, ranked_preys):
# same syntax as make_up_next_predator_generations, but with preys
p = [ranked_preys[i][0].brain for i in range(3)]
p += [ranked_preys[4][0].brain]
kids = []
for i in range(self.card_prey_pool - 2):
w = np.zeros(p[0].coef_.shape)
p1, p2 = np.random.choice(p, 2, replace=False)
for row in range(w.shape[0]):
for col in range(w.shape[1]):
r = np.random.random()
if r < .45: w[row, col] = p1.coef_[row, col]
elif r < .9: w[row, col] = p2.coef_[row, col]
else: w[row, col] = np.random.random() * 2 - 1
kid = Prey(Grid(*self.grid_size))
kid.brain.coef_ = w
kids.append(kid)
return kids, p[0], p[1]
def __init__(self, Environment, location):
self.Environment = Environment
# Initialize Q
alpha = 0.3
gamma = 0.7
epsilon = 0.1
self.Prey = Prey(Environment, location)
self.actions = Prey.actions
self.TeamQLearning = TeamQLearning(self, alpha, gamma, epsilon)
def __init__(self,
width=11,
height=11,
preyLocation=(5, 5),
predatorLocation=(0, 0)):
self.width = width
self.height = height
self.S, self.terminal_states = self.getStates()
self.Prey = Prey(self, preyLocation)
self.Predators = Predator(self, predatorLocation)
def __init__( self,
width=11,
height=11,
preyLocation=(5,5),
predatorLocation=(0,0),
numberOfPredators=1 ):
assert numberOfPredators > 0
assert numberOfPredators < 5
assert type(numberOfPredators) == int
self.width = width
self.height = height
self.numberOfPredators = numberOfPredators
self.Prey = Prey( self, preyLocation )
self.PredatorLocations = [(0,0), (10,0), (0,10), (10,10)]
self.Predators = [Predator(self, self.PredatorLocations[i]) \
for i in range(self.numberOfPredators)]
self.Agents = [self.Prey] + self.Predators
def __init__(self, width, height, color, grid):
Prey.__init__(self, width, height, color, grid)
self.offspringsProtected = 0
PreyAdult.dictionaryOfPreyAdults[(self.gridX, self.gridY)] = self
def __init__(self, width, height, color, grid):
Prey.__init__(self, width, height, color, grid)
self.offspringsProtected = 0
PreyAdult.dictionaryOfPreyAdults[(self.gridX, self.gridY)] = self
class TeamPrey():
def __init__(self, Environment, location):
self.Environment = Environment
# Initialize Q
alpha = 0.3
gamma = 0.7
epsilon = 0.1
self.Prey = Prey(Environment, location)
self.actions = Prey.actions
self.TeamQLearning = TeamQLearning(self, alpha, gamma, epsilon)
def updateQ(self, s, a, o, s_prime, r):
'''
Update this teams Q and V.
'''
O = self.actions
A = self.actions
# Use linear programming to obtain optimal policy for this state
try:
# Create a new model
m = grb.Model("MultiAgentMinimax")
m.setParam("OutputFlag", 0)
# Create variables
pi = dict()
for a in A:
pi[a] = m.addVar(0.0,
1.0,
vtype=grb.GRB.CONTINUOUS,
name=str(a))
# Integrate new variables
m.update()
# Set objective
m.setObjective(
grb.LinExpr([(self.TeamQLearning.Q[s][(a, o)], pi[a])
for o in O for a in A]), grb.GRB.MAXIMIZE)
# Add constraint: Sum_a pi(a) = 1
expr = grb.quicksum(m.getVars())
m.addConstr(expr == 1, "Total probability")
# Add more constraints
for o in O:
expr = grb.LinExpr([(self.TeamQLearning.Q[s][(a, o)], pi[a])
for a in A])
m.addConstr(expr >= 0)
m.optimize()
for a in A:
self.TeamQLearning.policy[s][a] = pi[a].x
except grb.GurobiError:
print 'Error reported'
# Update Q and V
self.TeamQLearning.updateQ(s, a, o, s_prime, r)
def getActionEpsilonGreedy(self, s):
# Find the (joint) action that maximizes Q[(s, a)]
prob_actions = dict()
uniform_epsilon = self.TeamQLearning.epsilon / (len(self.actions))
for possible_a in self.actions:
# Set probabilities of all actions uniformly
prob_actions[possible_a] = uniform_epsilon
best_a = argmax(self.TeamQLearning.policy[s])
prob_actions[best_a] += 1 - self.TeamQLearning.epsilon
# For every action, check if the cumulative probability exceeds a
# random number.
random_number = random.random()
cumulative_prob = 0.0
for a in self.actions:
cumulative_prob += prob_actions[a]
if cumulative_prob >= random_number:
return a
def performAction(self, a):
self.Prey.performAction(a)
def permutations(self, iterable, r=None):
'''
iterator <- permutations(iterable, r)
Finds permutations of iterable of length r, with duplicate entries.
'''
pool = tuple(iterable)
n = len(pool)
r = n if r is None else r
for indices in product(range(n), repeat=r):
if len(indices) == r:
yield tuple(pool[i] for i in indices)
def generate_random_preys(self, n):
return [Prey(Grid(*self.grid_size)) for _ in range(n)]
class TeamPrey():
def __init__(self, Environment, location):
self.Environment = Environment
# Initialize Q
alpha = 0.3
gamma = 0.7
epsilon = 0.1
self.Prey = Prey(Environment, location)
self.actions = Prey.actions
self.TeamQLearning = TeamQLearning(self, alpha, gamma, epsilon)
def updateQ(self, s, a, o, s_prime, r):
'''
Update this teams Q and V.
'''
O = self.actions
A = self.actions
# Use linear programming to obtain optimal policy for this state
try:
# Create a new model
m = grb.Model("MultiAgentMinimax")
m.setParam("OutputFlag",0)
# Create variables
pi = dict()
for a in A:
pi[a] = m.addVar( 0.0, 1.0, vtype = grb.GRB.CONTINUOUS, name = str(a) )
# Integrate new variables
m.update()
# Set objective
m.setObjective( grb.LinExpr( [ ( self.TeamQLearning.Q[s][(a,o)], pi[a] ) for o in O for a in A ] ), grb.GRB.MAXIMIZE)
# Add constraint: Sum_a pi(a) = 1
expr = grb.quicksum( m.getVars() )
m.addConstr( expr == 1, "Total probability" )
# Add more constraints
for o in O:
expr = grb.LinExpr( [ (self.TeamQLearning.Q[s][(a,o)], pi[a]) for a in A ] )
m.addConstr( expr >= 0 )
m.optimize()
for a in A:
self.TeamQLearning.policy[s][a] = pi[a].x
except grb.GurobiError:
print 'Error reported'
# Update Q and V
self.TeamQLearning.updateQ(s, a, o, s_prime, r)
def getActionEpsilonGreedy(self, s):
# Find the (joint) action that maximizes Q[(s, a)]
prob_actions = dict()
uniform_epsilon = self.TeamQLearning.epsilon / (len(self.actions))
for possible_a in self.actions:
# Set probabilities of all actions uniformly
prob_actions[possible_a] = uniform_epsilon
best_a = argmax( self.TeamQLearning.policy[s] )
prob_actions[best_a] += 1 - self.TeamQLearning.epsilon
# For every action, check if the cumulative probability exceeds a
# random number.
random_number = random.random()
cumulative_prob = 0.0
for a in self.actions:
cumulative_prob += prob_actions[a]
if cumulative_prob >= random_number:
return a
def performAction(self, a):
self.Prey.performAction(a)
def permutations(self, iterable, r=None):
'''
iterator <- permutations(iterable, r)
Finds permutations of iterable of length r, with duplicate entries.
'''
pool = tuple(iterable)
n = len(pool)
r = n if r is None else r
for indices in product(range(n), repeat=r):
if len(indices) == r:
yield tuple(pool[i] for i in indices)
def visualize(evolution, n_predator, n_prey, prey_card):
predator = evolution.predator_pool[n_predator]
prey = evolution.prey_pool[n_prey]
preys = [Prey(predator.grid, brain=prey.brain) for i in range(prey_card)]
predator.grid.visualize()
print('There are', predator.grid.n_preys(), 'preys left alive.\n')
def __init__(self,width,height,color,grid):
Prey.__init__(self, width, height, color, grid)
PreyOffspring.dictionaryOfOffsprings[(self.gridX,self.gridY)] = self
sizeOfScreen = numOfGridsInARow * sizeOfGrid
surface = pygame.display.set_mode((numOfGridsInARow * sizeOfGrid,numOfGridsInARow * sizeOfGrid))
numOfPreys = 1
numOfPredators = 1
numOfObstacles = 1
numOfFood = 1
worldMap = Map(numOfGridsInARow,sizeOfGrid,WHITE,surface,pygame)
# fixed item should be placed first, since our grid is implemented in a stack manner
obstacles = [Obstacle(worldMap,random.randrange(1,4),random.randrange(1,4),BROWN) for i in range(numOfObstacles)]
# obstacles = [Obstacle(worldMap,3,1,BROWN) for i in range(numOfObstacles)]
foods = [Food(worldMap,1,1,GREEN) for i in range(numOfFood)]
predators = [Predator(worldMap,3,1,1,RED) for i in range(numOfPredators)]
preys = [Prey(worldMap,2,1,1,YELLOW) for i in range(numOfPreys)]
age = 1
learningAges = 1
preysEaten = []
foodEaten = []
while True:
event = pygame.event.poll()
if event.type == pygame.QUIT:
pygame.quit()
break
surface.fill(WHITE)
#update for every animat in the world
# eatF = 0
