def e_greedy_search()

in leaderboard/llvm_instcount/e_greedy/e_greedy.py [0:0]

• 42 lines of code
• 12 McCabe index (conditional complexity)

def e_greedy_search(env: LlvmEnv) -> None:
    """Run an ϵ-greedy search on an environment."""
    step_count = 0
    with ThreadPoolExecutor(max_workers=FLAGS.nproc) as executor:
        while True:
            step_count += 1

            if random.random() < FLAGS.epsilon:
                # Exploratory step. Randomly select and apply an action.
                action = env.action_space.sample()
                _, reward, done, _ = env.step(action)
                logging.debug(
                    "Step %d, exploratory action %s, reward %.4f, cumulative %.4f",
                    step_count,
                    env.action_space.flags[action],
                    reward,
                    env.episode_reward,
                )
            else:
                # Select the best reward and apply it, or terminate the search
                # if no positive reward is attainable.
                best = select_best_action(env, executor)
                if best.reward <= 0:
                    logging.debug(
                        "Greedy search terminated after %d steps, "
                        "no further reward attainable",
                        step_count,
                    )
                    done = True
                else:
                    _, reward, done, _ = env.step(best.action)
                    logging.debug(
                        "Step %d, greedy action %s, reward %.4f, cumulative %.4f",
                        step_count,
                        env.action_space.flags[best.action],
                        reward,
                        env.episode_reward,
                    )
                    if env.reward_space.deterministic and reward != best.reward:
                        logging.warning(
                            "Action %s produced different reward on replay, %.4f != %.4f",
                            env.action_space.flags[best.action],
                            best.reward,
                            reward,
                        )

                # Stop the search if we have reached a terminal state.
                if done:
                    return