예제 #1
0
def sim_fixed(muA,
              muB,
              sample_size,
              n_experiments,
              sim_data=None,
              relative_mde_value=0.02,
              toc_adj_factor=0.00025,
              fn=None):

    if sim_data is None:
        sim_data = helper.sim_binomial_seq([muA, muB],
                                           sample_size=sample_size,
                                           n_experiments=n_experiments)

    el_res = ['U'] * n_experiments
    el_lift = [0] * n_experiments
    for s in range(n_experiments):
        sa = sim_data[s][0][sample_size - 1]
        sb = sim_data[s][1][sample_size - 1]
        res = el_core.calc_stat(sa, sample_size, sb, sample_size,
                                relative_mde_value, toc_adj_factor)
        el_res[s] = res['el_res']
        el_lift[s] = res['lift']

    res = pd.Series(el_res).value_counts().to_dict()

    helper.fprint(
        "[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
        " relative_mde_value: {}, toc: {:.10f}".format(muA, muB, sample_size,
                                                       n_experiments,
                                                       relative_mde_value,
                                                       toc), fn)

    winner_lift = []
    if muA == muB:

        for i, r in enumerate(el_res):
            if r in (['A', 'B']):
                winner_lift.append(el_lift[i])

        helper.fprint(
            "[EL] Fixed - Type 1 error rate: {:.4f}, error_avg_lift: {:.5f}, el_res: {}"
            .format(1 - (res.get('U', 0) + res.get('E', 0)) / n_experiments,
                    np.mean(abs(np.array(winner_lift))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(el_res):
            if r == winner:
                winner_lift.append(el_lift[i])

        helper.fprint(
            "[EL] Fixed - Power: {:.4f}, avg_lift: {:.5f}, el_res: {}".format(
                res.get(winner, 0) / n_experiments,
                np.mean(abs(np.array(winner_lift))), res), fn)

    return {'el_res': el_res, 'el_lift': el_lift}
예제 #2
0
def sim_fixed(muA, muB, sample_size, n_experiments, sim_data=None, alpha=0.05, fn=None):

    if sim_data is None:
        sim_data = helper.sim_binomial_seq([muA, muB], sample_size=sample_size, n_experiments=n_experiments, 
                                           cumsum=False)

    ttest_res = ['U']*n_experiments
    ttest_effect = [0]*n_experiments
    for s in range(n_experiments):
        sample_a = sim_data[s][0]
        sample_b = sim_data[s][1]
        ret = ttest_core.calc_stat(sample_a, sample_b, alpha)
        ttest_effect[s] = ret['effect']

        if (ret['p_value'] < alpha):
            if ret['cvr_a'] > ret['cvr_b']:
                ttest_res[s] = 'A'
            else:
                ttest_res[s] = 'B'

    helper.fprint("[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,}"
                  .format(muA, muB, sample_size, n_experiments), fn)

    winner_effect = []
    res = pd.Series(ttest_res).value_counts().to_dict()

    if muA == muB:

        for i, r in enumerate(ttest_res):
            if r in (['A','B']):
                winner_effect.append(ttest_effect[i])

        helper.fprint("[T-Test] Fixed - Type 1 error rate: {:.4f}, error_avg_effect: {:.5f}, ttest_res: {}"
                      .format(1-(res.get('U',0)+res.get('E',0))/n_experiments,
                              np.mean(abs(np.array(winner_effect))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(ttest_res):
            if r == winner:
                winner_effect.append(ttest_effect[i])

        helper.fprint("[T-Test] Fixed - Power: {:.4f}, avg_effect: {:.5f}, ttest_res: {}"
                      .format(res.get(winner,0)/n_experiments, np.mean(abs(np.array(winner_effect))), res), fn)

    return {'ttest_res': ttest_res,
            'ttest_effect': ttest_effect}
def sim_fixed(muA, muB, sample_size, n_experiments, sim_data=None, relative_mde_value=0.02, alpha=0.05):

    if sim_data is None:
        sim_data = helper.sim_binomial_seq([muA, muB], sample_size=sample_size, n_experiments=n_experiments)

    msprt_res = ['U']*n_experiments
    msprt_effect = [0]*n_experiments
    for s in range(n_experiments):
        sa = sim_data[s][0][sample_size-1]
        sb = sim_data[s][1][sample_size-1]
        res = msprt_core.calc_stat(sa, sample_size, sb, sample_size, relative_mde_value, alpha, 
                                   tau_option, tau_constant) 
        msprt_res[s] = res['msprt_res']
        msprt_effect[s] = res['effect']

    res = pd.Series(msprt_res).value_counts().to_dict()

    absolute_mde_value = muA*relative_mde_value
    min_msprt_half_width = absolute_mde_value*msprt_mde_ratio

    helper.fprint("[Parameters]: muA:{}, muB:{}, sample_size:{:,}, n_experiments:{:,},"
                  " min_msprt_half_width:{:.6f}"
                  .format(muA, muB, sample_size, n_experiments, relative_mde_value), fn)

    winner_effect = []
    if muA == muB:

        for i, r in enumerate(msprt_res):
            if r in (['A','B']):
                winner_effect.append(msprt_effect[i])

        helper.fprint("[mSPRT] Fixed - Type 1 error rate:{:.4f}, error_avg_effect:{:.5f}, msprt_res:{}"
                      .format(1-(res.get('U',0)+res.get('E',0))/n_experiments,
                              np.mean(abs(np.array(winner_effect))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(msprt_res):
            if r == winner:
                winner_effect.append(msprt_effect[i])

        helper.fprint("[mSPRT] Fixed - Power:{:.4f}, avg_effect:{:.5f}, msprt_res:{}"
                      .format(res.get(winner,0)/n_experiments, np.mean(abs(np.array(winner_effect))), res), fn)

    return {'msprt_res': msprt_res, 
            'msprt_effect': msprt_effect}
예제 #4
0
def sim_peeking(muA, muB, sample_size, n_experiments, n_peeks=-1, start=None, 
                alpha=0.05, relative_mde_value=0.02, 
                burnIn=300, random_seed=-1,
                pr_peek=1, pr_fixed=False, fn=None):

    if random_seed > 0:
        np.random.seed(random_seed)

    start_time = time.time()
    helper.fprint("[Start time]: {}".format(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    ss = helper.sample_size(muA, relative_mde_value=relative_mde_value)
    freq_sample_size = ss['variant_sample_size']

    if isinstance(start, float):
        assert (0 < start and start < 1)
        start = int(freq_sample_size*start)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size-start)/(n_peeks-1))
    elif isinstance(start, int):
        assert (1 <= start and start <= sample_size)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size-start)/(n_peeks-1))
    else:
        if n_peeks == -1:
            start = 1
            step = 1
        else:
            start = int(sample_size/n_peeks)
            step = start

    # print parameter settings
    helper.fprint("[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
                  " relative_mde_value:{:.5f}, alpha: {:.3f}, burnIn:{:,}, random_seed:{},"
                  " n_peeks: {:,}, start: {:,}, step: {:,}"
                  .format(muA, muB, sample_size, n_experiments, 
                          relative_mde_value, alpha, burnIn, random_seed,
                          n_peeks, start, step), fn)

    sim_data = helper.sim_binomial_seq([muA, muB], sample_size=sample_size, n_experiments=n_experiments, 
                                       cumsum=False)

    sim_res_data = []
    ttest_res = ['U']*n_experiments
    ttest_ss = [0]*n_experiments
    ttest_effect = [0]*n_experiments
    peek_idx = 0
    for n in range(start, sample_size+1, step):  
        peek_idx += 1
        peek_effect = []
        for s in range(n_experiments):
            if ttest_res[s] == 'U':
                sample_a = sim_data[s][0][:n]
                sample_b = sim_data[s][1][:n]
                ret = ttest_core.calc_stat(sample_a, sample_b, alpha, burnIn)
                cvr_a = ret['cvr_a']    
                cvr_b = ret['cvr_b']
                effect = ret['effect']
                ttest_res[s] = ret['ttest_res']
                ttest_ss[s] = n
                ttest_effect[s] = effect
                peek_effect.append(effect)
            
                if (ret['p_value'] < alpha):
                    if cvr_a > cvr_b:
                        ttest_res[s] = 'A'
                    else:
                        ttest_res[s] = 'B'
                    peek_effect.append(effect)

        res = pd.Series(ttest_res).value_counts().to_dict()
        peek_avg_effect = np.mean(abs(np.array(peek_effect)))

        sim_res_data_dict = {}
        sim_res_data_dict.update(res)
        sim_res_data_dict['peek_id'] = peek_idx
        sim_res_data_dict['samples'] = n
        sim_res_data_dict['avg_effect'] = peek_avg_effect
        sim_res_data.append(sim_res_data_dict)

        # print peek summary
        if pr_peek > 0 and peek_idx%pr_peek == 0:
            helper.fprint("[{}]: Peek #{} @ {:,} samples, avg_effect: {:.5f}, ttest_res: {}"
                          .format(time.strftime("%m-%d %H:%M:%S", time.localtime()), 
                                  peek_idx, n, peek_avg_effect, res), fn)
                
        if 'U' not in ttest_res:
            break

    # print parameter settings
    helper.fprint("[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
                  " relative_mde_value:{:.5f}, alpha: {:.3f}, burnIn:{:,}, random_seed:{},"
                  " n_peeks: {:,}, start: {:,}, step: {:,}"
                  .format(muA, muB, sample_size, n_experiments, 
                          relative_mde_value, alpha, burnIn, random_seed,
                          n_peeks, start, step), fn)


    # print simulation summary
    winner_ss = []
    winner_effect = []
    if muA == muB:

        for i, r in enumerate(ttest_res):
            if r in (['A','B']):
                winner_effect.append(ttest_effect[i])

        helper.fprint("[T-Test] {:,} peeks - Type 1 error rate: {:.4f},"
                      "error_avg_effect: {:.5f}, ttest_res: {}"
                      .format(peek_idx, 1-(res.get('U',0)+res.get('E',0))/n_experiments,
                              np.mean(abs(np.array(winner_effect))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(ttest_res):
            if r == winner:
                winner_ss.append(ttest_ss[i])
                winner_effect.append(ttest_effect[i])

        helper.fprint("[T-Test] {:,} peeks - Power: {:.4f}, avg_sample_size: {:,.0f}, sample_size_ratio: {:.2f}%,"
                      " avg_effect: {:.5f}, ttest_res: {}"
                      .format(peek_idx, res.get(winner,0)/n_experiments, np.mean(winner_ss),
                              np.mean(winner_ss)/freq_sample_size*100, 
                              np.mean(abs(np.array(winner_effect))), res), fn)

    if pr_fixed:
        sim_fixed(muA, muB, sample_size, n_experiments, sim_data, alpha) 

    helper.fprint("[Elasped time]: {}".format(str(timedelta(seconds=time.time()-start_time))), fn)
    helper.fprint("[End time]: {}".format(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    return {'sim_res_df': pd.DataFrame(sim_res_data),
            'ttest_res': ttest_res,
            'ttest_effect': ttest_effect}
예제 #5
0
def sim_peeking(muA,
                muB,
                sample_size,
                n_experiments,
                n_peeks=-1,
                start=None,
                relative_mde_value=0.02,
                rope_hdi_ratio=0.6,
                rope_mde_ratio=0.375,
                hdi_cred_mass=0.95,
                rvs_size=1000000,
                pr_peeking=True,
                pr_fixed=False,
                fn=None):

    start_time = time.time()
    helper.fprint(
        "[Start time]: {}".format(
            time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    ss = helper.sample_size(muA,
                            mde_value=relative_mde_value,
                            relative_mde=True)
    freq_sample_size = ss['variant_sample_size']

    if isinstance(start, float):
        assert (0 < start and start < 1)
        start = int(freq_sample_size * start)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size - start) / (n_peeks - 1))
    elif isinstance(start, int):
        assert (1 <= start and start <= sample_size)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size - start) / (n_peeks - 1))
    else:
        if n_peeks == -1:
            start = 1
            step = 1
        else:
            start = int(sample_size / n_peeks)
            step = start

    absolute_mde_value = muA * relative_mde_value
    min_rope_half_width = absolute_mde_value * rope_mde_ratio

    helper.fprint(
        "[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
        " relative_mde_value: {}, rope_hdi_ratio: {:.2f}, rope_mde_ratio: {:.3f},"
        " min_rope_half_width: {:.5f}, n_peeks: {:,}, start: {:,}, step: {:,}, rvs_size: {:,}"
        .format(muA, muB, sample_size, n_experiments, relative_mde_value,
                rope_hdi_ratio, rope_mde_ratio, min_rope_half_width, n_peeks,
                start, step, rvs_size), fn)

    sim_data = helper.sim_binomial_seq([muA, muB],
                                       sample_size=sample_size,
                                       n_experiments=n_experiments)

    sim_res_data = []
    rope_res = ['U'] * n_experiments
    rope_ss = [0] * n_experiments  # record the sample size when concluded
    rope_lift = [0] * n_experiments
    peek_idx = 0
    for n in range(start, sample_size + 1, step):
        peek_idx += 1
        peek_hdi_width = []
        peek_rope_width = []
        peek_lift = []
        for s in range(n_experiments):
            if rope_res[s] == 'U':
                sa = sim_data[s][0][n - 1]
                sb = sim_data[s][1][n - 1]
                res = rope_core.calc_stat(sa, n, sb, n, relative_mde_value,
                                          rope_hdi_ratio, rope_mde_ratio,
                                          hdi_cred_mass, rvs_size)
                rope_res[s] = res['rope_res']
                rope_ss[s] = n
                rope_lift[s] = res['lift']
                peek_hdi_width.append(res['lift_hdi_width'])
                peek_rope_width.append(res['rope_width'])
                if res['rope_res'] != 'U':
                    peek_lift.append(res['lift'])

        res = pd.Series(rope_res).value_counts().to_dict()
        peek_avg_hdi_width = np.mean(peek_hdi_width)
        peek_avg_rope_width = np.mean(peek_rope_width)
        peek_avg_lift = np.mean(abs(np.array(peek_lift)))

        sim_res_data_dict = {}
        sim_res_data_dict.update(res)
        sim_res_data_dict['peek_id'] = peek_idx
        sim_res_data_dict['samples'] = n
        sim_res_data_dict['avg_hdi_width'] = peek_avg_hdi_width
        sim_res_data_dict['avg_rope_width'] = peek_avg_rope_width
        sim_res_data_dict['avg_lift'] = peek_avg_lift
        sim_res_data.append(sim_res_data_dict)

        # Summary for each peek
        if pr_peeking:
            helper.fprint(
                "[{}]: Peek #{:,} @ {:,} samples, avg_hdi_width: {:.5f}, avg_rope_width: {:.5f},"
                " avg_lift: {:.5f}, rope_res: {}".format(
                    time.strftime("%m-%d %H:%M:%S", time.localtime()),
                    peek_idx, n, peek_avg_hdi_width, peek_avg_rope_width,
                    peek_avg_lift, res), fn)

        if 'U' not in rope_res:
            break

    # Simulation summary
    helper.fprint(
        "[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
        " relative_mde_value: {},"
        " rope_hdi_ratio: {:.2f}, rope_mde_ratio: {:.3f}, min_rope_half_width: {:.5f},"
        " n_peeks: {:,}, start: {:,}, step: {:,}, rvs_size: {:,}".format(
            muA, muB, sample_size, n_experiments, relative_mde_value,
            rope_hdi_ratio, rope_mde_ratio, min_rope_half_width, n_peeks,
            start, step, rvs_size), fn)

    winner_ss = []
    winner_lift = []
    if muA == muB:

        for i, r in enumerate(rope_res):
            if r in (['A', 'B']):
                winner_lift.append(rope_lift[i])

        helper.fprint(
            "[ROPE] {:,} peeks - Type 1 error rate: {:.4f}, error_avg_lift: {:.5f}, rope_res: {}"
            .format(peek_idx,
                    1 - (res.get('U', 0) + res.get('E', 0)) / n_experiments,
                    np.mean(abs(np.array(winner_lift))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(rope_res):
            if r == winner:
                winner_ss.append(rope_ss[i])
                winner_lift.append(rope_lift[i])

        helper.fprint(
            "[ROPE] {:,} peeks - Power: {:.4f}, avg_sample_size: {:,.0f},"
            " sample_size_ratio: {:.2f}%,"
            " avg_lift: {:.5f}, rope_res: {}".format(
                peek_idx,
                res.get(winner, 0) / n_experiments, np.mean(winner_ss),
                np.mean(winner_ss) / freq_sample_size * 100,
                np.mean(abs(np.array(winner_lift))), res), fn)

    # Show fixed sample size result
    if pr_fixed:
        sim_fixed(muA, muB, sample_size, n_experiments, sim_data,
                  relative_mde_value, rope_hdi_ratio, rope_mde_ratio,
                  hdi_cred_mass, rvs_size, fn)

    helper.fprint(
        "[Elasped time]: {}".format(
            str(timedelta(seconds=time.time() - start_time))), fn)
    helper.fprint(
        "[End time]: {}".format(
            time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    return {
        'sim_res_df': pd.DataFrame(sim_res_data),
        'rope_res': rope_res,
        'rope_lift': rope_lift
    }
예제 #6
0
def sim_fixed(muA,
              muB,
              sample_size,
              n_experiments,
              sim_data=None,
              relative_mde_value=0.02,
              rope_hdi_ratio=0.6,
              rope_mde_ratio=0.375,
              hdi_cred_mass=0.95,
              rvs_size=1000000,
              fn=None):

    if sim_data is None:
        sim_data = helper.sim_binomial_seq([muA, muB],
                                           sample_size=sample_size,
                                           n_experiments=n_experiments)

    rope_res = ['U'] * n_experiments
    rope_lift = [0] * n_experiments
    for s in range(n_experiments):
        sa = sim_data[s][0][sample_size - 1]
        sb = sim_data[s][1][sample_size - 1]
        res = rope_core.calc_stat(sa, sample_size, sb, sample_size,
                                  relative_mde_value, rope_hdi_ratio,
                                  rope_mde_ratio, hdi_cred_mass, rvs_size)
        rope_res[s] = res['rope_res']
        rope_lift[s] = res['lift']

    res = pd.Series(rope_res).value_counts().to_dict()

    absolute_mde_value = muA * relative_mde_value
    min_rope_half_width = absolute_mde_value * rope_mde_ratio

    helper.fprint(
        "[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
        " relative_mde_value: {}, rope_hdi_ratio: {}, rope_mde_ratio: {},"
        " min_rope_half_width: {:.6f}, rvs_size: {:,}".format(
            muA, muB, sample_size, n_experiments, relative_mde_value,
            rope_hdi_ratio, rope_mde_ratio, min_rope_half_width, rvs_size), fn)

    winner_lift = []
    if muA == muB:

        for i, r in enumerate(rope_res):
            if r in (['A', 'B']):
                winner_lift.append(rope_lift[i])

        helper.fprint(
            "[ROPE] Fixed - Type 1 error rate: {:.4f}, error_avg_lift: {:.5f}, rope_res: {}"
            .format(1 - (res.get('U', 0) + res.get('E', 0)) / n_experiments,
                    np.mean(abs(np.array(winner_lift))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(rope_res):
            if r == winner:
                winner_lift.append(rope_lift[i])

        helper.fprint(
            "[ROPE] Fixed - Power: {:.4f}, avg_lift: {:.5f}, rope_res: {}".
            format(
                res.get(winner, 0) / n_experiments,
                np.mean(abs(np.array(winner_lift))), res), fn)

    return {'rope_res': rope_res, 'rope_lift': rope_lift}
예제 #7
0
def sim_peeking(muA,
                muB,
                sample_size,
                n_experiments,
                n_peeks=-1,
                start=None,
                alpha=0.05,
                relative_mde_value=0.02,
                tau_option=3,
                tau_constant=0.0001,
                burnIn=0,
                random_seed=-1,
                pr_peek=1,
                pr_fixed=False,
                fn=None):

    if random_seed > 0:
        np.random.seed(random_seed)

    start_time = time.time()
    helper.fprint(
        "[Start time]:{}".format(
            time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    ss = helper.sample_size(muA, relative_mde_value=relative_mde_value)
    freq_sample_size = ss['variant_sample_size']

    if isinstance(start, float):
        assert (0 < start and start < 1)
        start = int(freq_sample_size * start)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size - start) / (n_peeks - 1))
    elif isinstance(start, int):
        assert (1 <= start and start <= sample_size)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size - start) / (n_peeks - 1))
    else:
        if n_peeks == -1:
            start = 1
            step = 1
        else:
            start = int(sample_size / n_peeks)
            step = start

    absolute_mde_value = muA * relative_mde_value

    # print parameter settings
    helper.fprint(
        "[Parameters]: muA:{}, muB:{}, sample_size:{:,}, n_experiments:{:,},"
        " relative_mde_value:{:.5f}, alpha:{:.3f}, tau_option:{}, tau_constant:{},"
        " burnIn:{:,}, random_seed:{}, n_peeks:{:,}, start:{:,}, step: {:,}".
        format(muA, muB, sample_size, n_experiments, relative_mde_value, alpha,
               tau_option, tau_constant, burnIn, random_seed, n_peeks, start,
               step), fn)

    sim_data = helper.sim_binomial_seq([muA, muB],
                                       sample_size=sample_size,
                                       n_experiments=n_experiments)

    sim_res_data = []
    msprt_res = ['U'] * n_experiments
    msprt_ss = [0] * n_experiments  # record the sample size when concluded
    msprt_effect = [0] * n_experiments
    msprt_prev_p_val = [1] * n_experiments
    msprt_prev_ci = [[-1000, 1000]] * n_experiments
    peek_idx = 0
    for n in range(start, sample_size + 1, step):
        peek_idx += 1
        peek_sig_effect = []
        peek_effect = []
        peek_decision_boundary = []
        for s in range(n_experiments):
            if msprt_res[s] == 'U':
                sa = sim_data[s][0][n - 1]
                sb = sim_data[s][1][n - 1]
                res = msprt_core.calc_stat(sa, n, sb, n, alpha,
                                           msprt_prev_p_val[s],
                                           msprt_prev_ci[s])
                msprt_res[s] = res['msprt_res']
                msprt_ss[s] = n
                msprt_effect[s] = res['effect']
                msprt_prev_p_val[s] = res['p_value']
                msprt_prev_ci[s] = res['effect_ci']
                if res['msprt_res'] != 'U':
                    peek_sig_effect.append(res['effect'])
                peek_effect.append(res['effect'])
                peek_decision_boundary.append(res['decision_boundary'])

        res = pd.Series(msprt_res).value_counts().to_dict()
        if len(peek_sig_effect) == 0:
            peek_avg_sig_effect = 0
        else:
            peek_avg_sig_effect = np.mean(abs(np.array(peek_sig_effect)))

        if len(peek_effect) == 0:
            peek_avg_effect = 0
        else:
            peek_avg_effect = np.mean(abs(np.array(peek_effect)))

        if len(peek_decision_boundary) == 0:
            peek_avg_decision_boundary = 0
        else:
            peek_avg_decision_boundary = np.mean(
                abs(np.array(peek_decision_boundary)))

        sim_res_data_dict = {}
        sim_res_data_dict.update(res)
        sim_res_data_dict['peek_id'] = peek_idx
        sim_res_data_dict['samples'] = n
        sim_res_data_dict['decision_boundary'] = peek_avg_decision_boundary
        sim_res_data_dict['avg_sig_effect'] = peek_avg_sig_effect
        sim_res_data_dict['avg_effect'] = peek_avg_effect
        sim_res_data.append(sim_res_data_dict)

        # peek summary
        if pr_peek > 0 and peek_idx % pr_peek == 0:
            helper.fprint(
                "[{}]: Peek #{:,} @ {:,} samples,"
                " avg_sig_effect:{:.5f},"
                " avg_effect:{:.5f}, msprt_res:{}".format(
                    time.strftime("%m-%d %H:%M:%S", time.localtime()),
                    peek_idx, n, peek_avg_sig_effect, peek_avg_effect, res),
                fn)

        if 'U' not in msprt_res:
            break

    # print parameters settings
    helper.fprint(
        "[Parameters]: muA:{}, muB:{}, sample_size:{:,}, n_experiments:{:,},"
        " relative_mde_value:{:.5f}, alpha:{:.3f}, tau_option:{}, tau_constant:{},"
        " burnIn:{:,}, random_seed:{}, n_peeks:{:,}, start:{:,}, step:{:,}".
        format(muA, muB, sample_size, n_experiments, relative_mde_value, alpha,
               tau_option, tau_constant, burnIn, random_seed, n_peeks, start,
               step), fn)

    # print simulation summary
    winner_ss = []
    winner_effect = []
    if muA == muB:

        for i, r in enumerate(msprt_res):
            if r in (['A', 'B']):
                winner_effect.append(msprt_effect[i])

        helper.fprint(
            "[mSPRT][{:,} peeks][{:,} samples] - Type 1 error rate:{:.4f},"
            " error_avg_effect:{:.5f}, msprt_res:{}".format(
                peek_idx, n,
                1 - (res.get('U', 0) + res.get('E', 0)) / n_experiments,
                np.mean(abs(np.array(winner_effect))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(msprt_res):
            if r == winner:
                winner_ss.append(msprt_ss[i])
                winner_effect.append(msprt_effect[i])

        helper.fprint(
            "[mSPRT][{:,} peeks][{:,} samples] - Power:{:.4f}, avg_sample_size:{:,.0f},"
            " sample_size_ratio:{:.2f}%, avg_effect:{:.5f}, msprt_res:{}".
            format(peek_idx, n,
                   res.get(winner, 0) / n_experiments, np.mean(winner_ss),
                   np.mean(winner_ss) / freq_sample_size * 100,
                   np.mean(abs(np.array(winner_effect))), res), fn)

    # show fixed sample size result
    if pr_fixed:
        sim_fixed(muA, muB, sample_size, n_experiments, sim_data,
                  relative_mde_value, alpha)

    helper.fprint(
        "[Elasped time]:{}".format(
            str(timedelta(seconds=time.time() - start_time))), fn)
    helper.fprint(
        "[End time]:{}".format(
            time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    return {
        'sim_res_df': pd.DataFrame(sim_res_data),
        'msprt_res': msprt_res,
        'msprt_effect': msprt_effect
    }
예제 #8
0
def sim_peeking(muA,
                muB,
                sample_size,
                n_experiments,
                n_peeks=-1,
                start=None,
                relative_mde_value=0.02,
                toc_adj_factor=0.00025,
                pr_peeking=True,
                pr_fixed=False,
                fn=None):

    start_time = time.time()
    helper.fprint(
        "[Start time]: {}".format(
            time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    ss = helper.sample_size(muA,
                            mde_value=relative_mde_value,
                            relative_mde=True)
    freq_sample_size = ss['variant_sample_size']

    if isinstance(start, float):
        assert (0 < start and start < 1)
        start = int(freq_sample_size * start)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size - start) / (n_peeks - 1))
    elif isinstance(start, int):
        assert (1 <= start and start <= sample_size)
        if n_peeks == -1:
            step = 1
        else:
            step = int((sample_size - start) / (n_peeks - 1))
    else:
        if n_peeks == -1:
            start = 1
            step = 1
        else:
            start = int(sample_size / n_peeks)
            step = start

    helper.fprint(
        "[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
        " relative_mde_value: {}, toc_adj_factor: {:.5f}, n_peeks: {:,}, start: {:,}, step: {:,}"
        .format(muA, muB, sample_size, n_experiments, relative_mde_value,
                toc_adj_factor, n_peeks, start, step), fn)

    sim_data = helper.sim_binomial_seq([muA, muB],
                                       sample_size=sample_size,
                                       n_experiments=n_experiments)

    sim_res_data = []
    el_res = ['U'] * n_experiments
    el_ss = [0] * n_experiments
    el_lift = [0] * n_experiments
    peek_idx = 0
    for n in range(start, sample_size + 1, step):
        peek_idx += 1
        peek_lift = []
        for s in range(n_experiments):
            if el_res[s] == 'U':
                sa = sim_data[s][0][n - 1]
                sb = sim_data[s][1][n - 1]
                res = el_core.calc_stat(sa, n, sb, n, relative_mde_value,
                                        toc_adj_factor)
                el_res[s] = res['el_res']
                el_ss[s] = n
                el_lift[s] = res['lift']
                if res['el_res'] != 'U':
                    peek_lift.append(res['lift'])

        res = pd.Series(el_res).value_counts().to_dict()
        peek_avg_lift = np.mean(abs(np.array(peek_lift)))

        sim_res_data_dict = {}
        sim_res_data_dict.update(res)
        sim_res_data_dict['peek_id'] = peek_idx
        sim_res_data_dict['samples'] = n
        sim_res_data_dict['avg_lift'] = peek_avg_lift
        sim_res_data.append(sim_res_data_dict)

        helper.fprint(
            "[{}]: Peek #{:,} @ {:,} samples, avg_lift: {:.5f}, el_res: {}".
            format(time.strftime("%m-%d %H:%M:%S", time.localtime()), peek_idx,
                   n, peek_avg_lift, res), fn)

        if ('U' not in el_res):
            break

    # Summary
    helper.fprint(
        "[Parameters]: muA: {}, muB: {}, sample_size: {:,}, n_experiments: {:,},"
        " relative_mde_value: {}, toc_adj_factor: {:.5f}, n_peeks: {:,}, start: {:,}, step: {:,}"
        .format(muA, muB, sample_size, n_experiments, relative_mde_value,
                toc_adj_factor, n_peeks, start, step), fn)

    winner_ss = []
    winner_lift = []
    if muA == muB:

        for i, r in enumerate(el_res):
            if r in (['A', 'B']):
                winner_lift.append(el_lift[i])

        if pr_peeking:
            helper.fprint(
                "[EL] {:,} peeks - Type 1 error rate: {:.4f}, error_avg_lift: {:.5f}, el_res: {}"
                .format(
                    peek_idx,
                    1 - (res.get('U', 0) + res.get('E', 0)) / n_experiments,
                    np.mean(abs(np.array(winner_lift))), res), fn)
    else:
        if muA > muB:
            winner = 'A'
        else:
            winner = 'B'

        for i, r in enumerate(el_res):
            if r == winner:
                winner_ss.append(el_ss[i])
                winner_lift.append(el_lift[i])

        if pr_peeking:
            helper.fprint(
                "[EL] {:,} peeks - Power: {:.4f}, avg_sample_size: {:,.0f},"
                "sample_size_ratio: {:.2f}%, avg_lift: {:.5f}, el_res: {}".
                format(peek_idx,
                       res.get(winner, 0) / n_experiments, np.mean(winner_ss),
                       np.mean(winner_ss) / freq_sample_size * 100,
                       np.mean(abs(np.array(winner_lift))), res), fn)

    if pr_fixed:
        sim_fixed(muA, muB, sample_size, n_experiments, sim_data,
                  relative_mde_value)

    helper.fprint(
        "[Elasped time]: {}".format(
            str(timedelta(seconds=time.time() - start_time))), fn)
    helper.fprint(
        "[End time]: {}".format(
            time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())), fn)

    return {
        'sim_res_df': pd.DataFrame(sim_res_data),
        'el_res': el_res,
        'el_lift': el_lift
    }