def find_failure_case(num_bases, num_blocks, max_levels):
while True:
# print("trying..")
env = BlocksWorldEnv(show=False)
thing_below, goal_thing_below = random_problem_instance(
env, num_blocks, max_levels, num_bases)
am = make_abstract_machine(env, num_bases, max_levels)
am_results = run_machine(am, goal_thing_below, {"jnt": "rest"})
env.close()
ticks, running_time, sym_reward, spa_reward = am_results
if sym_reward <= -2: break
# print(sym_reward)
return thing_below, goal_thing_below, sym_reward
def run_trial(num_bases, num_blocks, max_levels):
env = BlocksWorldEnv(show=False)
# rejection sample non-trivial instance
thing_below, goal_thing_below = random_problem_instance(
env, num_blocks, max_levels, num_bases)
am = make_abstract_machine(env, num_bases, max_levels)
nvm = virtualize(am)
am_results = run_machine(am, goal_thing_below, {"jnt": "rest"})
env.reset()
env.load_blocks(thing_below, num_bases)
nvm_results = run_machine(nvm, goal_thing_below,
{"jnt": tr.tensor(am.ik["rest"]).float()})
env.close()
return am_results, nvm_results, nvm.size(), thing_below, goal_thing_below
# init baseline
baseline = 0
for epoch in range(num_epochs):
start_epoch = time.perf_counter()
epoch_rewards = []
epoch_baselines = []
epoch_rtgs = []
for episode in range(num_episodes):
start_episode = time.perf_counter()
# random problem instance
env = BlocksWorldEnv(
show=showenv, step_hook=penalty_tracker.step_hook)
thing_below, goal_thing_below = random_problem_instance(
env, num_blocks, max_levels, num_bases)
nvm.env = env
reward, log_prob, rewards, log_probs = run_episode(
env, thing_below, goal_thing_below, nvm, init_regs,
init_conns, penalty_tracker, sigma)
env.close()
if use_penalties:
rewards = np.array(rewards)
rewards_to_go = np.cumsum(rewards)
rewards_to_go = rewards_to_go[
-1] - rewards_to_go + rewards
if reward <= reject_above:
for t in range(len(rewards)):