def main():
# 이미지 저장 경로 확인 및 생성
if not os.path.exists('images/'):
os.makedirs('images/')
# 그리드 월드 환경 객체 생성
env = GridWorld(
height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=None, # exploring start
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
# n-스텝
n = 3
# 스텝 사이즈
alpha = 0.2
# 가 수행당 1번의 에피소드 수행
episodes = 1000
values = dict()
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
values[(i, j)] = 0.0
for ep in range(episodes):
temporal_difference(env, values, n, alpha)
draw_grid_world_image(values, 'images/grid_world_fixed_params.png',
GRID_HEIGHT, GRID_WIDTH)
def main():
# 이미지 저장 경로 확인 및 생성
if not os.path.exists('images/'):
os.makedirs('images/')
# 그리드 월드 환경 객체 생성
env = GridWorld(height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=(0, 0),
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
values, returns = first_visit_mc_prediction(env, 1.0, 10000)
print("First Visit")
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print("({0}, {1}): {2:5.2f}".format(i, j, values[i, j]))
print()
draw_grid_world_image(values, 'images/first_visit_mc_state_values.png',
GRID_HEIGHT, GRID_WIDTH)
print()
values, returns = every_visit_mc_prediction(env, 1.0, 10000)
print("Every Visit")
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print("({0}, {1}): {2:5.2f}".format(i, j, values[i, j]))
print()
draw_grid_world_image(values, 'images/every_visit_mc_state_values.png',
GRID_HEIGHT, GRID_WIDTH)
def main():
# 그리드 월드 환경 객체 생성
env = GridWorld(
height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=None, # exploring start
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
MC = OffPolicyMonteCarloControl(env)
MC.off_policy_control()
print_grid_world_policy(env, MC.target_policy)
def grid_world_policy_evaluation():
# 그리드 월드 환경 객체 생성
env = GridWorld(height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=(0, 0),
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
# 수렴 시킨 상태 가치를 이미지로 저장하고 반복 횟수 반환 받음
state_values, iteration = compute_state_value(env)
print('정책 평가 --> 상태 가치 수렴: {} 회 반복'.format(iteration))
print(state_values)
draw_grid_world_image(np.round(state_values, decimals=2),
'images/state_values.png', GRID_HEIGHT, GRID_WIDTH)
def main():
# 그리드 월드 환경 객체 생성
env = GridWorld(
height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=None, # exploring start
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
MC = SoftPolicyMonteCarloControl(env)
MC.soft_policy_control()
with np.printoptions(precision=2, suppress=True):
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print(
i, j, ": UP, DOWN, LEFT, RIGHT", MC.policy[(i, j)][1],
env.action_space.ACTION_SYMBOLS[np.argmax(
MC.target_policy[(i, j)][1])])
print()
def main():
# 그리드 월드 환경 객체 생성
env = GridWorld(height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=(0, 0),
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
MC = OffPolicyMonteCarloPrediction(env)
MC.off_policy_prediction()
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print("({0}, {1}):".format(i, j))
for action in range(NUM_ACTIONS):
print(" Action {0}: {1:5.2f}".format(
env.action_space.ACTION_SYMBOLS[action],
MC.state_action_values[((i, j), action)]))
print()
print()
def main():
# 그리드 월드 환경 객체 생성
env = GridWorld(height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=(0, 0),
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0)
env.reset()
state_action_values, returns = first_visit_mc_prediction(env, 1.0, 10000)
print("First Visit")
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print("({0}, {1}):".format(i, j))
for action in range(NUM_ACTIONS):
print(" Action {0}: {1:5.2f}".format(
env.action_space.ACTION_SYMBOLS[action],
state_action_values[((i, j), action)]))
print()
print()
state_action_values, returns = every_visit_mc_prediction(env, 1.0, 10000)
print("Every Visit")
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print("({0}, {1}):".format(i, j))
for action in range(NUM_ACTIONS):
print(" Action {0}: {1:5.2f}".format(
env.action_space.ACTION_SYMBOLS[action],
state_action_values[((i, j), action)]))
print()
print()
iter_num_policy_evaluation = self.value_evaluation()
print("*** 가치 평가 [수렴까지 누적 반복 횟수: {0}] ***".format(iter_num_policy_evaluation))
self.policy_setup()
print("*** 정책 셋업 완료 ***")
return self.state_values, self.policy
if __name__ == '__main__':
# 그리드 월드 환경 객체 생성
env = GridWorld(
height=GRID_HEIGHT,
width=GRID_WIDTH,
start_state=(0, 0),
terminal_states=TERMINAL_STATES,
transition_reward=-1.0,
terminal_reward=-1.0,
outward_reward=-1.0
)
env.reset()
VI = ValueIteration(env)
VI.start_iteration()
print(VI.state_values, end="\n\n")
for i in range(GRID_HEIGHT):
for j in range(GRID_WIDTH):
print(i, j, VI.policy[i][j]) # UP, DOWN, LEFT, RIGHT
print()