baselines.py

# rlte/baselines.py
# Zwei Benchmark-Algorithmen für die Auswertung (Sec. 3.4): SL und TWAP.
# DR (Dirichlet-RL) ist im Paper nur Vergleich; hier bewusst NICHT implementiert (kein alternativer RL-Ansatz).
# :contentReference[oaicite:8]{index=8}
from __future__ import annotations
from typing import Dict, Tuple, List
import numpy as np

from .env_lob import LOBExecutionEnv
from . import config as C

def run_SL(market: str, M0: int, episodes: int = 1000) -> Tuple[float, float]:
    """
    Submit & Leave: zu t=0 ganze Menge M0 als Limit-Sell zum Best-Ask (Sec. 3.4), Rest bei T per Market.
    """
    rewards = []
    for ep in range(episodes):
        env = LOBExecutionEnv(market=market, M0=M0, seed=C.SEED + 1000 + ep)
        s = env.reset()
        # Aktion: a_K+1 simplex – alles auf a_k mit k="1 Tick über Best-Bid" (≈ Best-Ask bei Spread 1)
        K = C.K_SIMPLEX
        a = np.zeros(K + 1, dtype=np.float32)
        # Best-Ask entspricht (bei Spread≈1) k=1; wir setzen fast alles auf a1, Rest Hold=0
        a[1] = 1.0
        for n in range(C.N_STEPS):
            s, r, done, info = env.step(a)
            if done:
                break
            # danach "halten"
            a = np.zeros(K + 1, dtype=np.float32)
            a[-1] = 1.0
        rewards.append(r if 'r' in locals() else 0.0)
    return float(np.mean(rewards)), float(np.std(rewards))

def run_TWAP(market: str, M0: int, episodes: int = 1000) -> Tuple[float, float]:
    """
    TWAP: teile M0 in N Blöcke à M0/N; zu jedem Schritt Limit-Sell M0/N zum Best-Ask, Rest am Ende Market. (Sec. 3.4)
    """
    rewards = []
    block = int(round(M0 / C.N_STEPS))
    for ep in range(episodes):
        env = LOBExecutionEnv(market=market, M0=M0, seed=C.SEED + 2000 + ep)
        s = env.reset()
        K = C.K_SIMPLEX
        a = np.zeros(K + 1, dtype=np.float32)
        # Anteil block/M(t): Wir approximieren mit Simplex-Gewichten
        for n in range(C.N_STEPS):
            M = max(1, env.M_remaining)
            frac = min(1.0, block / float(M))
            a[:] = 0.0
            a[1] = frac       # Limit am Best-Ask (≈ k=1)
            a[-1] = 1.0 - frac
            s, r, done, info = env.step(a)
            if done:
                break
        rewards.append(r if 'r' in locals() else 0.0)
    return float(np.mean(rewards)), float(np.std(rewards))