def cost_benchmark():
sample_state = MiniSoccerState.generate_start_state()
player = sample_state.index['player']
opponent = sample_state.index['opponent']
# player_on_left = sample_state.index['player_on_left']
player_has_ball = sample_state.index['player_has_ball']
right_goal_center = sample_state.index['rightgoalcenter']
left_goal_center = sample_state.index['leftgoalcenter']
upper_left = sample_state.index['upperleft']
print "training the base agent..."
base_features = [rl.FeatureFlag(player_has_ball),
rl.FeatureAngle(player, upper_left, left_goal_center),
rl.FeatureDistY(player, right_goal_center),
rl.FeaturePointXY(player)
]
base_agent = MiniSoccerAgent(rl.FeatureSet(base_features))
a = time.clock()
arbitrator = rl.ArbitratorStandard(base_agent, NUM_TRIALS, NUM_EPISODES)
arbitrator.run(MAX_STEPS)
b = time.clock()
base_time = b - a
print "Running time: %.1f" % base_time
print "Do it again..."
a = time.clock()
arbitrator = rl.ArbitratorStandard(base_agent, NUM_TRIALS, NUM_EPISODES)
arbitrator.run(MAX_STEPS)
b = time.clock()
base_time = b - a
print "Running time: %.1f" % base_time
feature_lists = [
[rl.FeatureFlag(player_has_ball)],
[rl.FeatureAngle(opponent, left_goal_center, upper_left)],
[rl.FeatureAngle(opponent, left_goal_center, upper_left, 20)],
[rl.FeatureDist(opponent, player)],
[rl.FeatureDist(opponent, player, 20)],
[rl.FeatureDistX(opponent, player)],
[rl.FeatureDistX(opponent, player, 20)],
[rl.FeaturePointXY(opponent)],
[rl.FeaturePointXY(opponent, 400)],
[rl.FeatureInteraction([rl.FeatureDist(opponent, player), rl.FeatureAngle(opponent, left_goal_center, upper_left)])],
[rl.FeatureInteraction([rl.FeatureDist(opponent, player, 20), rl.FeatureAngle(opponent, left_goal_center, upper_left)])],
[rl.FeatureInteraction([rl.FeatureDist(opponent, player), rl.FeaturePointXY(opponent)])],
[rl.FeatureInteraction([rl.FeatureDist(opponent, player, 20), rl.FeaturePointXY(opponent)])],
]
for feature_list in feature_lists:
agent = base_agent.clone()
for feature in feature_list:
agent.add_feature(feature)
arbitrator = rl.ArbitratorStandard(agent, NUM_TRIALS, NUM_EPISODES)
print "testing %s..." % feature_list
a = time.clock()
arbitrator.run(MAX_STEPS)
b = time.clock()
print "Overhead time: %.1f" % (b - a - base_time)
print
def learn_w_multitile_features():
sample_state = MiniSoccerState.generate_start_state()
player = sample_state.index['player']
opponent = sample_state.index['opponent']
# player_on_left = sample_state.index['player_on_left']
player_has_ball = sample_state.index['player_has_ball']
right_goal_center = sample_state.index['rightgoalcenter']
left_goal_center = sample_state.index['leftgoalcenter']
features = [
rl.FeatureDist('dist-player-opponent', player, opponent),
rl.FeatureDist('dist-player-rightgoalcenter', player,
right_goal_center),
rl.FeatureDist('dist-opponent-rightgoalcenter', opponent,
right_goal_center),
rl.FeatureDist('dist-player-leftgoalcenter', player, left_goal_center),
rl.FeatureDist('dist-opponent-leftgoalcenter', opponent,
left_goal_center),
rl.FeatureFlag('flag-has-ball', player_has_ball)
]
offsets = rl.TiledFeature.EVEN_OFFSETS
feature_list = []
for offset in offsets:
for i in range(len(features)):
the_feature = copy.deepcopy(features[i])
the_feature.offset = offset
feature_list.append(the_feature)
agent = MiniSoccerAgent(rl.FeatureSet(feature_list))
arbitrator = rl.ArbitratorStandard(agent, NUM_TRIALS, NUM_EPISODES)
arbitrator.execute(MAX_STEPS)
def learn_w_multitile_features():
sample_state = MiniSoccerState.generate_start_state()
player = sample_state.index['player']
opponent = sample_state.index['opponent']
# player_on_left = sample_state.index['player_on_left']
player_has_ball = sample_state.index['player_has_ball']
right_goal_center = sample_state.index['rightgoalcenter']
features = [
rl.FeatureFlag(player_has_ball),
rl.FeatureAngle(player, opponent, right_goal_center),
rl.FeatureDist(player, opponent)
]
feature_list = features
agent = MiniSoccerAgent(rl.FeatureSet(feature_list))
arbitrator = rl.ArbitratorStandard(agent, NUM_TRIALS, NUM_EPISODES)
arbitrator.execute(MAX_STEPS)
def learn_w_multitile_features():
sample_state = MiniSoccerState.generate_start_state()
player = sample_state.index['player']
opponent = sample_state.index['opponent']
player_has_ball = sample_state.index['player_has_ball']
right_goal_center = sample_state.index['rightgoalcenter']
left_goal_center = sample_state.index['leftgoalcenter']
right_goal_top = sample_state.index['rightgoaltop']
right_goal_bottom = sample_state.index['rightgoalbottom']
left_goal_top = sample_state.index['leftgoaltop']
left_goal_bottom = sample_state.index['leftgoalbottom']
features = [
rl.FeatureDist(player, opponent),
rl.FeatureDist(player, right_goal_center),
rl.FeatureDist(player, left_goal_center),
rl.FeatureDist(opponent, right_goal_center),
rl.FeatureDist(opponent, left_goal_center),
rl.FeatureAngle(player, opponent, right_goal_top),
rl.FeatureAngle(player, opponent, left_goal_bottom)
]
offsets = rl.TiledFeature.EVEN_OFFSETS
feature_list = []
feature_list.append(rl.FeatureFlag(player_has_ball))
for offset in offsets:
for i in range(len(features)):
the_feature = copy.deepcopy(features[i])
the_feature.offset = offset
feature_list.append(the_feature)
agent = MiniSoccerAgent(rl.FeatureSet(feature_list))
arbitrator = rl.ArbitratorStandard(agent, NUM_TRIALS, NUM_EPISODES)
arbitrator.run(MAX_STEPS)
from minisoccer import MiniSoccerState
import rl
sample_state = MiniSoccerState.generate_start_state()
player_has_ball = sample_state.index['player_has_ball']
feature = rl.FeatureFlag('flag-has-ball', player_has_ball)
print feature.encode_state(sample_state)
sample_state.index['player_has_ball'].truth = False
print feature.encode_state(sample_state)