2021 01 12 Klines 1 loss function test - WojciechMigda/TruRL GitHub Wiki

Parametry eksperymentu: 
Episodes: 100
max_episode_steps: 200
Memory capacity: 100000
GAMMA: 0.70000
NEPOCHS(20)
KBinsDiscretizer({
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {34, -0.020000, 0.020000},
    {10, 0.000000, 100.000000},
    {10, 0.000000, 100.000000},
    {10, 0.000000, 200.000000},})
Scaler({[-5.000000, 5.000000], [0, 10000]})
TsetliniClassifierBitwise({
        "threshold": 10000,
        "s": 4.000000,
        "number_of_regressor_clauses": 3200,
        "number_of_states": 127,
        "boost_true_positive_feedback": 1,
        "random_state": 1,
        "n_jobs": 6,
        "clause_output_tile_size": 16,
        "weighted": true,
        "loss_fn": "L1+2",
        "loss_fn_C1": 0.100000,
        "max_weight": 2147483647,
        "verbose": false
    })
Gym: <TimeLimit<WavyMarketEnv, gen_fn=07_klines Actions=[<Actions.HOLD: 0>, <Actions.BUY100: 1>, <Actions.SELL100: 2>]>>

Q=±5, T=10k

Celem tego eksperymentu było uzyskanie wglądu na ile zmiana loss function może poprawić stabilność samouczenia.

Wyjściowy eksperyment to log #2 z 2021-01-10 Klines initiative 1.

Jak widać, niestety mimo że można mówić o pewnej poprawie (L1+2(0.5)), to cały czas mamy do czynienia z dużym rozrzutem.

Box Plot

Dane

Lokalizacja: /experiments/2021-01-12_klines_loss_fn

Uzyte skrypty są wersjonowane w katalogu powyżej.

⚠️ **GitHub.com Fallback** ⚠️