def execute(self, grades, moduleDict, solutionDict):
grades.addMessage('Testing Deep Q Network...')
# Load Pacman Agent
nographics = False
pacmanType = loadAgent("PacmanDeepQAgent", nographics)
pacman = pacmanType(self.layout)
# Load Ghost Agent
ghostType = loadAgent("RandomGhost", nographics)
numghosts = 1
ghosts = [ghostType(i + 1) for i in range(numghosts)]
numTraining = pacman.model.numTrainingGames # Set by student
numGames = numTraining + self.numEvalGames
record = False
games = runGames(self.layout,
self.horizon,
pacman,
ghosts,
self.display,
numGames,
record,
numTraining=numTraining,
catchExceptions=False,
timeout=30)
scores = [game.state.getScore() for game in games]
wins = [game.state.isWin() for game in games]
winRate = wins.count(True) / float(len(wins))
if winRate < self.winThresh:
grades.addMessage(
'FAIL:\nWinRate = {} < {} threshold for full credit'.format(
winRate, self.winThresh))
return False
elif winRate < self.winThreshEC:
grades.addMessage(
'PASS:\nWinRate = {} >= {} threshold for full credit'.format(
winRate, self.winThresh))
grades.assignFullCredit()
return True
else:
grades.addMessage(
'PASS:\nWinRate = {} >= {} threshold for extra credit'.format(
winRate, self.winThreshEC))
grades.assignFullCredit()
grades.addPoints(1)
return True
def eval_genomes_single(genomes, config):
for id, genome in genomes:
genome.fitness = 0
nn_model = neat.nn.FeedForwardNetwork.create(genome, config)
pacmanType = pacman.loadAgent("NEATAgent", True)
cmd_line_args['pacman'] = pacmanType(nn_model=nn_model)
cmd_line_args['display'] = textDisplay.NullGraphics()
#cmd_line_args['display'] = graphicsDisplay.PacmanGraphics()
games = pacman.runGames(**cmd_line_args)
for game in games:
genome.fitness += game.state.getScore()
def eval_genomes(genome, config):
genome.fitness = 0
nn_model = neat.nn.FeedForwardNetwork.create(genome, config)
pacmanType = pacman.loadAgent("NEATAgent", True)
cmd_line_args['pacman'] = pacmanType(nn_model=nn_model)
cmd_line_args['display'] = textDisplay.NullGraphics()
# cmd_line_args['display'] = graphicsDisplay.PacmanGraphics()
games = pacman.runGames(**cmd_line_args)
result = 0
for game in games:
result += game.state.getScore()
return float(result) / float(len(games))
def load_setup(options):
""" Load the setup for training the agent
"""
layout_ = layout.getLayout(options.layout)
if layout is None:
raise Exception("The layout " + options.layout + " cannot be found")
# Initialize a Pacman agent
pacmanType = loadAgent(options.pacman, nographics=False)
pacmanOpts = parseAgentArgs(options.agentArgs)
pacman = pacmanType(
**pacmanOpts
) # Instantiate Pacman with depth=2 # To Do: allow depth as an argument
# Initialize Ghost agents
ghostType = loadAgent(options.ghost, nographics=False)
ghostOpts = parseAgentArgs(options.ghostArgs)
ghosts = [
ghostType(index=i + 1, numTraining=options.numTraining, **ghostOpts)
for i in range(options.numGhosts)
]
return layout_, pacman, ghosts
def evaluate_candidates(candidates, args):
nn_model = args["nn_model"]
cmd_line_args = args["cmd_line_args"]
candidates_fitness = []
for candidate in candidates:
nn_model.set_weights(candidate)
pacmanType = pacman.loadAgent("BioAgent", True)
cmd_line_args['pacman'] = pacmanType(nn_model=nn_model)
cmd_line_args['display'] = textDisplay.NullGraphics()
#cmd_line_args['display'] = graphicsDisplay.PacmanGraphics()
games = pacman.runGames(**cmd_line_args)
candidate_fitness = 0
for game in games:
candidate_fitness += game.state.getScore()
candidates_fitness.append(candidate_fitness)
print(candidates_fitness)
return candidates_fitness
def load_ghosts(options, filename):
""" Load the ghosts from file
:param options: options from commandline
:param filename: pkl file that stores the agent
:return: ghosts set from pkl file
"""
# Initialize ghosts given type
ghostType = loadAgent(options.ghost, nographics=False)
ghostOpts = parseAgentArgs(options.ghostArgs)
ghosts = [
ghostType(index=i + 1, numTraining=options.numTraining, **ghostOpts)
for i in range(options.numGhosts)
]
# Load ghosts from file and turn off training
for ghost in ghosts:
ghost.turnoff_training()
ghost.load_from_file(filename)
return ghosts
# Add a stdout reporter to show progress in the terminal.
p.add_reporter(neat.StdOutReporter(True))
stats = neat.StatisticsReporter()
p.add_reporter(stats)
p.add_reporter(neat.Checkpointer(5))
# Run for up to 300 generations (parallel)
pe = neat.ParallelEvaluator(4, eval_genomes)
winner = p.run(pe.evaluate, 1000)
#winner = p.run(eval_genomes_single, 100)
# Display the winning genome.
print('\nBest genome:\n{!s}'.format(winner))
# Save the winner to disk
with open('winner-feedforward.out', 'wb') as f:
pickle.dump(winner, f)
# Show output of the most fit genome against training data.
print('\nOutput:')
winner_net = neat.nn.FeedForwardNetwork.create(winner, config)
# Load the final agent and run a few games with it
pacmanType = pacman.loadAgent("NEATAgent", True)
cmd_line_args['pacman'] = pacmanType(nn_model=winner_net)
cmd_line_args['display'] = textDisplay.NullGraphics()
#cmd_line_args['display'] = graphicsDisplay.PacmanGraphics()
cmd_line_args['numGames'] = 10
games = pacman.runGames(**cmd_line_args)
pass
def readCommand(argv):
"""
Processes the command used to run pacman from the command line.
"""
from optparse import OptionParser
usageStr = """
USAGE: python pacman.py <options>
EXAMPLES: (1) python pacman.py
- starts an interactive game
(2) python pacman.py --layout smallClassic --zoom 2
OR python pacman.py -l smallClassic -z 2
- starts an interactive game on a smaller board, zoomed in
"""
parser = OptionParser(usageStr)
parser.add_option('-n',
'--numGames',
dest='numGames',
type='int',
default=10,
help=default('the number of GAMES to play'),
metavar='GAMES')
parser.add_option('--numDisplay',
dest='numDisplay',
type='int',
default=10,
help=default('the number of GAMES to display'))
parser.add_option(
'-l',
'--layout',
dest='layout',
default='small',
help=default(
'the index of LAYOUT_FILE from which to load the map layout'))
parser.add_option(
'-p',
'--pacman',
dest='pacman',
default='MC',
help=default('the agent in the pacmanAgents module to use'))
parser.add_option('-a',
'--all',
dest='testAll',
action='store_true',
default=False,
help=default('if we want to test all the parameters'))
parser.add_option(
'-s',
'--showPlot',
dest='showPlot',
action='store_true',
default=False,
help=default('If we want to show the plot of the result'))
parser.add_option('--norun',
dest='noRun',
action='store_true',
default=False,
help=default('If run the game'))
parser.add_option('--savegif',
dest='savegif',
action='store_true',
default=False,
help=default('If save the gif files'))
options, otherjunk = parser.parse_args(argv)
if len(otherjunk) != 0:
raise Exception('Command line input not understood: ' + str(otherjunk))
args = dict()
ghostType = loadAgent('RandomGhost', False)
args['ghosts'] = [ghostType(i + 1) for i in range(10)]
args['numGames'] = options.numGames
args['numGamesToDisplay'] = options.numDisplay
argsList = []
if options.testAll:
for layout in layouts:
for pacman in pacmans.values():
argsTmp = copy.deepcopy(args)
argsTmp['layoutName'] = layout
argsTmp['pacman'] = pacman
argsList.append(argsTmp)
else:
args['layoutName'] = options.layout
args['pacman'] = pacmans[options.pacman]
argsList.append(args)
return (argsList, options)
if __name__ == '__main__':
cmd_line_args = pacman.readCommand(
sys.argv[1:]) # Get game components based on input
nn_model = generate_model()
prng = Random()
prng.seed(time())
ea = ec.DEA(prng)
ea.terminator = ec.terminators.evaluation_termination
final_pop = ea.evolve(generator=generate_candidate,
evaluator=evaluate_candidates,
pop_size=1000,
maximize=True,
max_evaluations=1000,
mutation_rate=0.3,
crossover_rate=1,
nn_model=nn_model,
num_elites=1,
mp_num_cpus=4,
cmd_line_args=cmd_line_args)
best = max(final_pop)
nn_model.set_weights(best.candidate)
print("Final candidate:", best.candidate)
pacmanType = pacman.loadAgent("BioAgent", True)
cmd_line_args['pacman'] = pacmanType(nn_model=nn_model)
cmd_line_args['display'] = graphicsDisplay.PacmanGraphics()
cmd_line_args['numGames'] = 10
games = pacman.runGames(**cmd_line_args)
pass
def read_command(argv):
from optparse import OptionParser
usage_str = """
USAGE: python experiment.py <options>
EXAMPLES: (1) python experiment.py --layout testRL --pacman=AlphaBetaAgent --ghosts=QLearningGhost --numTraining=10
"""
parser = OptionParser(usage_str)
parser.add_option(
'-l',
'--layout',
dest='layout',
help=default('the LAYOUT_FILE from which to load the map layout'),
metavar='LAYOUT_FILE',
default='mediumClassic')
parser.add_option(
'-p',
'--pacman',
dest='pacman',
help=default('the agent TYPE in the pacmanAgents module to use'),
metavar='TYPE',
default='KeyboardAgent')
parser.add_option(
'-a',
'--agentArgs',
dest='agentArgs',
help=
'Comma separated values sent to agent. e.g. "opt1=val1,opt2,opt3=val3"'
)
parser.add_option(
'-g',
'--ghosts',
dest='ghost',
help=default('the ghost agent TYPE in the ghostAgents module to use'),
metavar='TYPE',
default='QLearningGhost')
parser.add_option(
'-x',
'--numTraining',
dest='numTraining',
type='int',
help=default('How many episodes are training (suppresses output)'),
default=10)
parser.add_option('-k',
'--numghosts',
type='int',
dest='numGhosts',
help=default('The maximum number of ghosts to use'),
default=4)
parser.add_option(
'-f',
'--fixRandomSeed',
action='store_true',
dest='fixRandomSeed',
help='Fixes the random seed to always play the same game',
default=False)
parser.add_option(
'--ghostArgs',
dest='ghostArgs',
help=
'Comma separated values sent to ghost. e.g. "opt1=val1,opt2,opt3=val3"'
)
options, otherjunk = parser.parse_args(argv)
if len(otherjunk) != 0:
raise Exception('Command line input not understood: ' + str(otherjunk))
args = dict()
# Fix the random seed
if options.fixRandomSeed: random.seed('cs188')
# Choose a layout
args['layout'] = layout.getLayout(options.layout)
if args['layout'] is None:
raise Exception("The layout " + options.layout + " cannot be found")
# Initialize a Pacman agent
pacmanType = loadAgent(options.pacman, nographics=False)
pacmanOpts = parseAgentArgs(options.agentArgs)
pacman = pacmanType(
**pacmanOpts
) # Instantiate Pacman with depth=2 # To Do: allow depth as an argument
args['pacman'] = pacman
# Initialize Ghost agents
ghostType = loadAgent(options.ghost, nographics=False)
ghostOpts = parseAgentArgs(options.ghostArgs)
args['ghosts'] = [
ghostType(index=i + 1, numTraining=options.numTraining, **ghostOpts)
for i in range(options.numGhosts)
]
# Set number of training episodes for experiment
args['numTraining'] = options.numTraining
return args, options