sfMask = np.random.choice(a=[False, True], size=(16), p=[0.5, 0.5])
sfMask = np.concatenate([sfMask, [True]]) # constant term
sfTestMask = np.zeros(
shape=16,
dtype=np.bool) # State features rejected (we think the agent have)
sfTestTrials = np.zeros(
shape=16, dtype=np.int32) # Num of trials for each state feature
#
# Initial model - First test
#
print("Using initial MDP =\n", initialModel, flush=True)
mdp = gridworld.GridworldEnv(initialModel)
mdp.horizon = MDP_HORIZON
if SAVE_STATE_IMAGES:
saveStateImage("stateImage_" + str(experiment_i) + "_0A.png", mdp,
sfTestMask)
#agent_policy_initial_model = GaussianPolicy(nStateFeatures=np.count_nonzero(sfMask),actionDim=2)
#agent_learner_initial_model = GpomdpLearner(mdp,agent_policy_initial_model,gamma=0.98)
agent_policy_initial_model = BoltzmannPolicy(np.count_nonzero(sfMask), 4)
agent_learner_initial_model = GpomdpLearner(mdp,
agent_policy_initial_model,
gamma=0.98)
learn(learner=agent_learner_initial_model,
steps=LEARNING_STEPS,
nEpisodes=LEARNING_EPISODES,
import numpy as np
from gym.envs.toy_text import gridworld
from util.util_gridworld import *
from util.policy_boltzmann import *
from util.learner import *
np.set_printoptions(precision=6)
np.set_printoptions(suppress=True)
mdp = gridworld.GridworldEnv(changeProb=0.5, targetCol=2)
mdp.horizon = 50
sfMask = np.ones(shape=17, dtype=bool) # state features mask
sfMask[14] = False
# AGENT SPACE
agent_policy = BoltzmannPolicy(np.count_nonzero(sfMask), 4)
agent_learner = GpomdpLearner(mdp, agent_policy, gamma=0.98)
learn(agent_learner,
steps=0,
nEpisodes=1000,
sfmask=sfMask,
adamOptimizer=True,
learningRate=0.3,
loadFile=None,
saveFile=None,
autosave=False,
plotGradient=False)
# SUPERVISOR SPACE
import numpy as np
import sys
from gym.envs.toy_text import gridworld
from util.learner import *
from util.optimizer import *
from util.policy_boltzmann import *
from util.util_gridworld import *
w_row = [5, 1, 1, 1, 1]
w_col = [5, 1, 1, 1, 1]
w_grow = [5, 1, 1, 1, 1]
w_gcol = [1, 1, 1, 1, 5]
model = np.array([w_row, w_col, w_grow, w_gcol], dtype=np.float32)
mdp = gridworld.GridworldEnv(model)
mdp.horizon = 50
sfMask = np.array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1],
dtype=bool) # Agent features
sfTestMask = np.ones(shape=16, dtype=np.bool) # Features not rejected
#
# Cycle ENVIRONMENT CONFIGURATION
#
for conf_index in range(1000):
print("Using MDP ", model)
mdp = gridworld.GridworldEnv(model)
mdp.horizon = 50
saveStateImage("stateImage" + str(conf_index) + "A.png", mdp, sfTestMask)
import numpy as np
from gym.envs.toy_text import gridworld
from util.util_gridworld import *
from util.policy_boltzmann import *
from util.learner import *
np.set_printoptions(precision=6)
np.set_printoptions(suppress=True)
mdp = gridworld.GridworldEnv()
mdp.horizon = 50
sfMask = np.ones(shape=16,dtype=bool) # state features mask
sfMask[8:16] = False
# AGENT SPACE
agent_policy = BoltzmannPolicy(np.count_nonzero(sfMask),4)
agent_learner = GpomdpLearner(mdp,agent_policy,gamma=0.98)
learn(
agent_learner,
steps=30,
nEpisodes=250,
sfmask=sfMask,
adamOptimizer=True,
learningRate=0.3,
loadFile=None,#"params8.npy",
saveFile=None,
autosave=True,
plotGradient=False