def reward_fn(code_string):
return misc.RewardInfo(
episode_rewards=[float(ord(c)) for c in code_string],
input_case=[],
correct_output=[],
code_output=[],
input_type=misc.IOType.integer,
output_type=misc.IOType.integer,
reason='none')
def _score_string(self, string):
actions = misc.bf_string_to_tokens(string)
reward, correct = self.reward_fn(actions)
return misc.RewardInfo(episode_rewards=[0.0] * (len(string) - 1) +
[reward],
input_case=None,
correct_output=None,
code_output=actions,
input_type=None,
output_type=misc.IOType.integer,
reason='correct' if correct else 'wrong')
def _score_code(self, code):
"""Run test cases on code and compute reward.
Args:
code: A single BF code string.
Returns:
misc.RewardInfo namedtuple instance containing reward and code execution
information, including inputs, expected outputs, code outputs, input
and output types, and reason for the reward obtained.
"""
# Get list of 2-tuples, each containing an input sequence and an output
# sequence.
io_seqs = self.task.make_io_set()
terminal_reward = 0.0
results = []
reason = 'correct'
for input_seq, output_seq in io_seqs:
eval_result = bf.evaluate(
code, input_buffer=input_seq, timeout=0.1,
max_steps=self.max_execution_steps,
base=self.task.base,
require_correct_syntax=self.require_correct_syntax)
result, success = eval_result.output, eval_result.success
if not success:
# Code execution timed out.
terminal_reward = self.failure_reward
results = []
reason = eval_result.failure_reason
break
else:
terminal_reward += self.reward_fn(result, output_seq, self.task.base)
if result == output_seq:
terminal_reward += self.correct_bonus # Bonus for correct answer.
# Only add additional reward for shorter code. Subtracting reward
# interferes with the main objective. Only optimize for length once
# any solution is found.
if self.min_code_length == self.max_code_length:
terminal_reward += self.code_length_bonus
else:
terminal_reward += self.code_length_bonus * clipped_linear(
x=len(code), x0=self.min_code_length, y0=1.0,
slope=-self.time_penalty, y_range=(0.0, 1.0))
# reason remains 'correct' if it is already
elif reason == 'correct':
reason = 'wrong'
results.append(result)
# Return list of rewards, one for each char in the code. All are 0 except
# for the terminal reward.
terminal_reward /= self.best_reward
return misc.RewardInfo(
episode_rewards=[0.0] * (len(code) - 1) + [terminal_reward],
input_case=misc.IOTuple(i for i, o in io_seqs),
correct_output=misc.IOTuple(o for i, o in io_seqs),
code_output=misc.IOTuple(results),
input_type=self.input_type,
output_type=self.output_type,
reason=reason)
