Ejemplo n.º 1
0
 def simulate(trial_id):
     env = make_env('constant_high',
                    cost=COST,
                    ground_truth=TRIALS[trial_id])
     df = evaluate(pol, [env] * 30)
     df['trial_id'] = trial_id
     return df
Ejemplo n.º 2
0
def chunk_exp_Q(env_type, seed, chunk_i):
    name = f'{env_type}_{seed}'
    os.makedirs(f'data/exp_Q/{name}', exist_ok=True)
    polfile = f'data/policies/{name}.pkl'
    pol = load(polfile)
    chunk = STATE_CHUNKS[chunk_i]
    env = make_env('constant_high',
                   cost=COST,
                   term_belief=True,
                   ground_truth=False)

    def Q(state):
        for action in env.actions(state):
            if action == env.term_action:
                q = env.expected_term_reward(state)
            else:
                samples = []
                for _ in range(1000):
                    env._state = state
                    env.init, r, *_ = env.step(action)
                    samples.append(r + sum(run_episode(pol, env)['rewards']))
                q = np.mean(samples)
            yield {'state': state, 'action': action, 'q': q}

    pd.DataFrame(list(concat(map(
        Q, chunk)))).to_pickle(f'data/exp_Q/{name}/{chunk_i}.pkl')
Ejemplo n.º 3
0
def yoked_rollout(pol, trial_id, clicks, n=1):
    env = make_env('constant_high', cost=COST, ground_truth=TRIALS[trial_id])
    true_init = env.init

    for _ in range(n):
        env.init = true_init
        trace = run_episode(pol, env)
        yield {
            's': env.init,
            'a': trace['actions'][0],
            'q': sum(trace['rewards'])
        }

        for click in clicks:
            s = env.reset()
            env.init, r, *_ = env.step(click)
            q = r + sum(run_episode(pol, env)['rewards'])
            yield {'s': s, 'a': click, 'q': q}
Ejemplo n.º 4
0
def chunk_write_rollouts(env_type, seed, chunk_i):
    polfile = f'data/policies/{env_type}_{seed}.pkl'
    pol = load(polfile)
    env = make_env('constant_high', cost=COST)
    chunk = CLICK_CHUNKS[chunk_i]
    n = 0

    for i, (trial_id, clicks) in enumerate(chunk):
        data = {'s': [], 'a': [], 'q': [], 'phi': [], 'trial_id': trial_id}
        for step in yoked_rollout(pol, trial_id, clicks, n=300):
            data['s'].append(encode_state(step['s']))
            data['a'].append(step['a'])
            data['q'].append(step['q'])
            phi = pol.phi(step['s'], step['a'], compute_all=True)[1:]
            phi[0] = env.expected_term_reward(step['s'])
            data['phi'].append(phi)

        data['phi'] = np.stack(data['phi'])
        data['a'] = np.array(data['a'])
        data['q'] = np.array(data['q'])
        dump(data, f'data/rollouts/{env_type}_{seed}/{chunk_i}_{i}.pkl')
        n += len(data['q'])

    return n