2021 01 10 Klines 1 Q range - 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({[<#####>, <#####>], [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": "MSE",
"loss_fn_C1": 0.000000,
"max_weight": 2147483647,
"verbose": false
})
Gym: <TimeLimit<WavyMarketEnv, gen_fn=07_klines Actions=[<Actions.HOLD: 0>, <Actions.BUY100: 1>, <Actions.SELL100: 2>]>>
Zakresy funkcji Q wzięte pod uwagę: [-4, +4], [-5, +5], [-50, +50].
Wykonano 10 powtórzeń, po 100 epizodów, każdy epizod po 200 kroków.

Dla najszerszego zakresu (50) nie widać, żeby model czegoś się uczył. Znaczna poprawa nastąpiła dla Q +/-5, jeszcze większa dla +/-4.
Do sprawdzenia: 2, 3, 6.
Lokalizacja: /experiments/2021-01-10_klines_1m
Użyte skrypty są wersjonowane w katalogu powyżej.