2021 01 18 Klines 1 Threshold - WojciechMigda/TruRL GitHub Wiki

Parametry eksperymentu: 
Episodes: 100
max_episode_steps: 200
Memory capacity: 100000
GAMMA: 0.70000
LAMBDA: 0.00100
BATCH SIZE: 64
Agent's seed: 0
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({[-0.500000, 0.500000], [0, <#####>]})
TsetliniClassifierBitwise({
        "threshold": 5000,
        "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,
        "box_muller": false,
        "loss_fn": "MSE",
        "loss_fn_C1": 0.000000,
        "max_weight": 2147483647,
        "verbose": false
    })
Gym: <TimeLimit<WavyMarketEnv, gen_fn=07_klines (../../experiments/klines/Binance_BTCUSDT_1m_1597190400000-1598227200000.json) window_size=30 Actions=[<Actions.HOLD: 0>, <Actions.BUY100: 1>, <Actions.SELL100: 2>] Fees=0/0>>

Celem eksperymentu było porównanie jakości samouczenia dla różnych wartości Threshold. Wcześniejszy eksperyment był przeprowadzony dla innego zakresu roboczego funkcji Q.

Box Plot

Wydaje się, że T=10k wypada najlepiej.

Dane

Lokalizacja: /experiments/2021-01-18_klines_1m_Threshold

Użyte skrypty są wersjonowane w katalogu powyżej.

⚠️ **GitHub.com Fallback** ⚠️