def testAugmented():
from core import VGDLParser
from pybrain.rl.experiments.episodic import EpisodicExperiment
from mdpmap import MDPconverter
from agents import PolicyDrivenAgent
zelda_level2 = """
wwwwwwwwwwwww
wA wwk1ww w
ww ww 1 w
ww wwww+w
wwwww1ww www
wwwww 0 Gww
wwwwwwwwwwwww
"""
from examples.gridphysics.mazes.rigidzelda import rigidzelda_game
g = VGDLParser().parseGame(rigidzelda_game)
g.buildLevel(zelda_level2)
env = GameEnvironment(g,
visualize=False,
recordingEnabled=True,
actionDelay=150)
C = MDPconverter(g, env=env, verbose=True)
Ts, R, _ = C.convert()
print C.states
print Ts[0]
print R
env.reset()
agent = PolicyDrivenAgent.buildOptimal(env)
env.visualize = True
env.reset()
task = GameTask(env)
exper = EpisodicExperiment(task, agent)
exper.doEpisodes(1)
def testRecordingToGif(human=False):
from pybrain.rl.experiments.episodic import EpisodicExperiment
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_2
from agents import PolicyDrivenAgent, InteractiveAgent
from tools import makeGifVideo
game_str, map_str = polarmaze_game, maze_level_2
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g,
visualize=human,
recordingEnabled=True,
actionDelay=200)
task = GameTask(env)
if human:
agent = InteractiveAgent()
else:
agent = PolicyDrivenAgent.buildOptimal(env)
exper = EpisodicExperiment(task, agent)
res = exper.doEpisodes(1)
print res
actions = [a for _, a, _ in env._allEvents]
print actions
makeGifVideo(env, actions, initstate=env._initstate)
def playMultipleEpisodes(self, numEpisodes):
tally, finalEventList, totalStatesEncountered = [], [], []
gameObject = None
for i in range(numEpisodes):
print "Starting episode", i
self.pickNewLevel(index=i)
gameObject, won, eventList, statesEncountered = self.playEpisode(
gameObject, finalEventList)
finalEventList.extend(eventList)
VGDLParser.playGame(self.gameString,
self.levelString,
statesEncountered,
persist_movie=True,
make_images=True,
make_movie=False,
movie_dir="videos/" + self.gameName,
padding=10)
totalStatesEncountered.append(statesEncountered)
tally.append(won)
print "Episode ended. Won:", won
print "Have won", sum(tally), "out of", len(tally), "episodes"
print "Won", sum(tally), "out of ", len(tally), "episodes."
makeVideo(movie_dir="videos/" + self.gameName)
# empty image directory
# shutil.rmtree("images/tmp")
# os.makedirs("images/tmp")
return
def setup(name, mode):
if mode == "vgdl":
# Read in level and game file information
level = get_file(name, "level")
game = get_file(name, "game")
# Start game
g = VGDLParser().parseGame(game)
g.buildLevel(level)
rle = RLEnvironment(game, level, observationType='global', visualize=True)
# Set up RLE
rle.actionDelay = 200
rle.recordingEnabled = True
rle.reset()
rle._game._drawAll()
dims = rle.outdim
environment = rle
# TODO
elif mode == "ale":
return
else:
return
# Return environment
return environment, dims
def testAugmented():
from core import VGDLParser
from pybrain.rl.experiments.episodic import EpisodicExperiment
from mdpmap import MDPconverter
from agents import PolicyDrivenAgent
zelda_level2 = """
wwwwwwwwwwwww
wA wwk1ww w
ww ww 1 w
ww wwww+w
wwwww1ww www
wwwww 0 Gww
wwwwwwwwwwwww
"""
from examples.gridphysics.mazes.rigidzelda import rigidzelda_game
g = VGDLParser().parseGame(rigidzelda_game)
g.buildLevel(zelda_level2)
env = GameEnvironment(g, visualize=False,
recordingEnabled=True, actionDelay=150)
C = MDPconverter(g, env=env, verbose=True)
Ts, R, _ = C.convert()
print C.states
print Ts[0]
print R
env.reset()
agent = PolicyDrivenAgent.buildOptimal(env)
env.visualize = True
env.reset()
task = GameTask(env)
exper = EpisodicExperiment(task, agent)
exper.doEpisodes(1)
def testRollout(actions=[0, 0, 2, 2, 0, 3] * 20):
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from core import VGDLParser
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=True, actionDelay=100)
env.rollOut(actions)
def testRolloutVideo(actions=[0, 0, 2, 2, 0, 3] * 2):
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from core import VGDLParser
from tools import makeGifVideo
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
makeGifVideo(GameEnvironment(g, visualize=True), actions)
def testRollout(actions=[0, 0, 2, 2, 0, 3] * 20):
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from core import VGDLParser
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=True, actionDelay=100)
env.rollOut(actions)
def _createVGDLGame(gameSpec, levelSpec):
import uuid
from core import VGDLParser
# parse, run and play.
game = VGDLParser().parseGame(gameSpec)
game.buildLevel(levelSpec)
game.uiud = uuid.uuid4()
return game
def testRolloutVideo(actions=[0, 0, 2, 2, 0, 3] * 2):
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from core import VGDLParser
from tools import makeGifVideo
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
makeGifVideo(GameEnvironment(g, visualize=True), actions)
def testStochMaze():
from core import VGDLParser
from examples.gridphysics.mazes.stochastic import stoch_game, stoch_level
g = VGDLParser().parseGame(stoch_game)
g.buildLevel(stoch_level)
C = MDPconverter(g, verbose=True)
Ts, R, fMap = C.convert()
print C.states
print R
for T in Ts:
print T
print fMap
def testStochMaze():
from core import VGDLParser
from examples.gridphysics.mazes.stochastic import stoch_game, stoch_level
g = VGDLParser().parseGame(stoch_game)
g.buildLevel(stoch_level)
C = MDPconverter(g, verbose=True)
Ts, R, fMap = C.convert()
print C.states
print R
for T in Ts:
print T
print fMap
def test2():
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from core import VGDLParser
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
actions = [1, 0, 0, 3, 0, 2, 0, 2, 0, 0, 0]
env = GameEnvironment(g, visualize=True, actionDelay=100)
env.rollOut(actions)
env.reset()
senv = SubjectiveGame(g, actionDelay=1500)
senv.rollOut(actions)
def testMaze():
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
C = MDPconverter(g, verbose=True)
Ts, R, fMap = C.convert()
print C.states
print R
for T in Ts:
print T
print fMap
def test2():
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from core import VGDLParser
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
actions = [1, 0, 0, 3, 0, 2, 0, 2, 0, 0, 0]
env = GameEnvironment(g, visualize=True, actionDelay=100)
env.rollOut(actions)
env.reset()
senv = SubjectiveGame(g, actionDelay=1500)
senv.rollOut(actions)
def testMaze():
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
C = MDPconverter(g, verbose=True)
Ts, R, fMap = C.convert()
print C.states
print R
for T in Ts:
print T
print fMap
def testPolicyAgent():
from pybrain.rl.experiments.episodic import EpisodicExperiment
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_2
from agents import PolicyDrivenAgent
game_str, map_str = polarmaze_game, maze_level_2
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=False, actionDelay=100)
task = GameTask(env)
agent = PolicyDrivenAgent.buildOptimal(env)
env.visualize = True
env.reset()
exper = EpisodicExperiment(task, agent)
res = exper.doEpisodes(2)
print res
def testPolicyAgent():
from pybrain.rl.experiments.episodic import EpisodicExperiment
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_2
from agents import PolicyDrivenAgent
game_str, map_str = polarmaze_game, maze_level_2
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=False, actionDelay=100)
task = GameTask(env)
agent = PolicyDrivenAgent.buildOptimal(env)
env.visualize = True
env.reset()
exper = EpisodicExperiment(task, agent)
res = exper.doEpisodes(2)
print res
def test3():
from examples.gridphysics.mazes import polarmaze_game
from examples.gridphysics.mazes.simple import maze_level_1b
from core import VGDLParser
from pybrain.rl.experiments.episodic import EpisodicExperiment
from interfaces import GameTask
from agents import InteractiveAgent, UserTiredException
game_str, map_str = polarmaze_game, maze_level_1b
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
senv = SubjectiveGame(g, actionDelay=100, recordingEnabled=True)
#senv = GameEnvironment(g, actionDelay=100, recordingEnabled=True, visualize=True)
task = GameTask(senv)
iagent = InteractiveAgent()
exper = EpisodicExperiment(task, iagent)
try:
exper.doEpisodes(1)
except UserTiredException:
pass
print senv._allEvents
def test3():
from examples.gridphysics.mazes import polarmaze_game
from examples.gridphysics.mazes.simple import maze_level_1b
from core import VGDLParser
from pybrain.rl.experiments.episodic import EpisodicExperiment
from interfaces import GameTask
from agents import InteractiveAgent, UserTiredException
game_str, map_str = polarmaze_game, maze_level_1b
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
senv = SubjectiveGame(g, actionDelay=100, recordingEnabled=True)
# senv = GameEnvironment(g, actionDelay=100, recordingEnabled=True, visualize=True)
task = GameTask(senv)
iagent = InteractiveAgent()
exper = EpisodicExperiment(task, iagent)
try:
exper.doEpisodes(1)
except UserTiredException:
pass
print senv._allEvents
def testInteractions():
from random import randint
from pybrain.rl.experiments.episodic import EpisodicExperiment
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from pybrain.rl.agents.agent import Agent
class DummyAgent(Agent):
total = 4
def getAction(self):
res = randint(0, self.total - 1)
return res
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=True, actionDelay=100)
task = GameTask(env)
agent = DummyAgent()
exper = EpisodicExperiment(task, agent)
res = exper.doEpisodes(2)
print res
def testRecordingToGif(human=False):
from pybrain.rl.experiments.episodic import EpisodicExperiment
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_2
from agents import PolicyDrivenAgent, InteractiveAgent
from tools import makeGifVideo
game_str, map_str = polarmaze_game, maze_level_2
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=human, recordingEnabled=True, actionDelay=200)
task = GameTask(env)
if human:
agent = InteractiveAgent()
else:
agent = PolicyDrivenAgent.buildOptimal(env)
exper = EpisodicExperiment(task, agent)
res = exper.doEpisodes(1)
print res
actions = [a for _, a, _ in env._allEvents]
print actions
makeGifVideo(env, actions, initstate=env._initstate)
def testInteractions():
from random import randint
from pybrain.rl.experiments.episodic import EpisodicExperiment
from core import VGDLParser
from examples.gridphysics.mazes import polarmaze_game, maze_level_1
from pybrain.rl.agents.agent import Agent
class DummyAgent(Agent):
total = 4
def getAction(self):
res = randint(0, self.total - 1)
return res
game_str, map_str = polarmaze_game, maze_level_1
g = VGDLParser().parseGame(game_str)
g.buildLevel(map_str)
env = GameEnvironment(g, visualize=True, actionDelay=100)
task = GameTask(env)
agent = DummyAgent()
exper = EpisodicExperiment(task, agent)
res = exper.doEpisodes(2)
print res
def makeImages(gameName, gameString, levelString, statesEncountered, param_ID):
VGDLParser.playGame(gameString, levelString, statesEncountered, \
persist_movie=True, make_images=True, make_movie=False, movie_dir="videos/"+gameName, gameName = gameName, parameter_string=param_ID, padding=10)
# gameFilename = "examples.gridphysics.demo_multigoal_and" ##takes forever if you have many boxes and don't use 2BFS (with metabolic penalty) ## Continuous physics games can't work right now. RLE is discretized, getSensors() relies on this, and a lot of the induction/planning ## architecture depends on that. Will take some work to do this well. Best plan is to shrink the grid squares and increase speeds/strengths of ## objects. # gameFilename = "examples.continuousphysics.mario" # gameFilename = "examples.gridphysics.boulderdash" #Game is buggy. # gameFilename = "examples.gridphysics.butterflies" gameString, levelString = defInputGame(gameFilename, randomize=True) rleCreateFunc = lambda: createRLInputGame(gameFilename) rle = rleCreateFunc() p = IW(rle, gameString, levelString, gameFilename, k=2) VGDLParser.playGame(gameString, levelString, p.statesEncountered, persist_movie=True, make_images=True, make_movie=False, movie_dir="videos/"+gameFilename, padding=10) # embed() t1 = time.time() last, visited, rejected = BFS3(rle, p) print time.time()-t1 print len(visited), len(rejected) embed() #
