import argparse import datetime import json import math import random import re import subprocess import sys import threading PARAM_INFO = { "futilityMarginGain": {"range": (100, 200), "default": 165, "c_end": 20, "r_end": 0.0020}, "reductionA": {"range": (1000, 2000), "default": 1642, "c_end": 100, "r_end": 0.0020}, "reductionB": {"range": (500, 1500), "default": 1024, "c_end": 100, "r_end": 0.0020}, "reductionC": {"range": (500, 1500), "default": 916, "c_end": 100, "r_end": 0.0020}, "statBonusA": {"range": (1, 30), "default": 12, "c_end": 3, "r_end": 0.0020}, "statBonusB": {"range": (100, 500), "default": 282, "c_end": 40, "r_end": 0.0020}, "statBonusC": {"range": (100, 500), "default": 349, "c_end": 40, "r_end": 0.0020}, "statBonusD": {"range": (1000, 2000), "default": 1594, "c_end": 100, "r_end": 0.0020}, "reductionInit": {"range": (1000, 3000), "default": 2026, "c_end": 200, "r_end": 0.0020}, "counterMoveHistoryThreshold": { "range": (-150, 0), "default": -1, "c_end": 15, "r_end": 0.0020, }, "aspirationDeltaA": {"range": (0, 30), "default": 10, "c_end": 3, "r_end": 0.0020}, "aspirationDeltaB": { "range": (10000, 20000), "default": 15620, "c_end": 1000, "r_end": 0.0020, }, "aspirationDeltaC": {"range": (2, 7), "default": 4, "c_end": 0.5, "r_end": 0.0020}, "aspirationDeltaD": {"range": (1, 10), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "bonusInitialGain": {"range": (-100, 0), "default": -19, "c_end": 10, "r_end": 0.0020}, "bonusInitialThreshold": { "range": (1000, 3000), "default": 1914, "c_end": 200, "r_end": 0.0020, }, "improvementDefault": {"range": (0, 400), "default": 168, "c_end": 40, "r_end": 0.0020}, "mateBetaDelta": {"range": (50, 250), "default": 137, "c_end": 20, "r_end": 0.0020}, "mateDepthThreshold": {"range": (1, 10), "default": 5, "c_end": 0.5, "r_end": 0.0020}, "mateExtraBonus": {"range": (10, 100), "default": 62, "c_end": 9, "r_end": 0.0020}, "futilityBaseDelta": {"range": (50, 200), "default": 153, "c_end": 15, "r_end": 0.0020}, "razoringA": {"range": (-500, 0), "default": -369, "c_end": 50, "r_end": 0.0020}, "razoringB": {"range": (-500, 0), "default": -254, "c_end": 50, "r_end": 0.0020}, "futilityA": {"range": (100, 500), "default": 303, "c_end": 40, "r_end": 0.0020}, "futilityDepth": {"range": (3, 12), "default": 8, "c_end": 0.5, "r_end": 0.0020}, "nullMoveThreshA": { "range": (10000, 20000), "default": 17139, "c_end": 1000, "r_end": 0.0020, }, "nullMoveThreshB": {"range": (-100, 0), "default": -20, "c_end": 10, "r_end": 0.0020}, "nullMoveThreshC": {"range": (1, 20), "default": 13, "c_end": 1, "r_end": 0.0020}, "nullMoveThreshD": {"range": (100, 500), "default": 233, "c_end": 40, "r_end": 0.0020}, "nullMoveThreshE": {"range": (1, 50), "default": 25, "c_end": 5, "r_end": 0.0020}, "nullMoveRA": {"range": (100, 500), "default": 168, "c_end": 40, "r_end": 0.0020}, "nullMoveRB": {"range": (1, 20), "default": 7, "c_end": 1, "r_end": 0.0020}, "nullMoveRC": {"range": (1, 10), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "nullMoveRD": {"range": (1, 10), "default": 4, "c_end": 0.5, "r_end": 0.0020}, "nullMoveRE": {"range": (100, 1500), "default": 861, "c_end": 140, "r_end": 0.0020}, "nullMoveDepth": {"range": (8, 20), "default": 14, "c_end": 0.5, "r_end": 0.0020}, "nullMovePlyA": {"range": (1, 5), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "nullMovePlyB": {"range": (1, 8), "default": 4, "c_end": 0.8, "r_end": 0.0020}, "probCutBetaA": {"range": (100, 300), "default": 191, "c_end": 20, "r_end": 0.0020}, "probCutBetaB": {"range": (10, 100), "default": 54, "c_end": 9, "r_end": 0.0020}, "probCutDepthLimit": {"range": (1, 10), "default": 4, "c_end": 0.5, "r_end": 0.0020}, "probCutDepth": {"range": (1, 10), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "ttDecreaseA": {"range": (1, 5), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "ttDecreaseB": {"range": (1, 5), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "ttDecreaseDepth": {"range": (1, 20), "default": 9, "c_end": 0.5, "r_end": 0.0020}, "probCutBetaC": {"range": (300, 500), "default": 417, "c_end": 20, "r_end": 0.0020}, "probCutDepthThresh": {"range": (1, 5), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "shallowPruningDepthA": {"range": (1, 15), "default": 7, "c_end": 1, "r_end": 0.0020}, "shallowPruningA": {"range": (100, 300), "default": 180, "c_end": 20, "r_end": 0.0020}, "shallowPruningB": {"range": (100, 300), "default": 201, "c_end": 20, "r_end": 0.0020}, "shallowPruningC": {"range": (1, 10), "default": 6, "c_end": 0.5, "r_end": 0.0020}, "sseThreshold": {"range": (-500, 0), "default": -222, "c_end": 50, "r_end": 0.0020}, "shallowPruningDepthB": {"range": (1, 10), "default": 5, "c_end": 0.5, "r_end": 0.0020}, "shallowPruningD": { "range": (-5000, -1000), "default": -3875, "c_end": 400, "r_end": 0.0020, }, "shallowPruningGain": {"range": (1, 5), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "shallowPruningDepthC": {"range": (1, 20), "default": 13, "c_end": 1, "r_end": 0.0020}, "shallowPruningE": {"range": (50, 200), "default": 106, "c_end": 15, "r_end": 0.0020}, "shallowPruningF": {"range": (100, 200), "default": 145, "c_end": 10, "r_end": 0.0020}, "shallowPruningG": {"range": (10, 100), "default": 52, "c_end": 9, "r_end": 0.0020}, "shallowPruningH": {"range": (-100, 0), "default": -24, "c_end": 10, "r_end": 0.0020}, "shallowPruningI": {"range": (-50, 0), "default": -15, "c_end": 5, "r_end": 0.0020}, "singularExtDepthA": {"range": (1, 8), "default": 4, "c_end": 0.5, "r_end": 0.0020}, "singularExtDepthB": {"range": (1, 5), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "singularExtDepthC": {"range": (1, 5), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "singularBetaA": {"range": (1, 5), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "singularExtentionA": {"range": (10, 50), "default": 25, "c_end": 4, "r_end": 0.0020}, "singularExtentionB": {"range": (5, 15), "default": 9, "c_end": 0.5, "r_end": 0.0020}, "singularExtDepthD": {"range": (5, 15), "default": 9, "c_end": 0.5, "r_end": 0.0020}, "singularExtentionC": {"range": (50, 150), "default": 82, "c_end": 10, "r_end": 0.0020}, "singularExtentionD": { "range": (1000, 10000), "default": 5177, "c_end": 900, "r_end": 0.0020, }, "lmrDepthThreshold": {"range": (1, 5), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "lmrMoveCountThreshold": {"range": (1, 15), "default": 7, "c_end": 0.5, "r_end": 0.0020}, "lmrDecTTPv": {"range": (0, 3), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "lmrDecMoveCount": {"range": (0, 3), "default": 1, "c_end": 0.5, "r_end": 0.0020}, "lmrDecSingular": {"range": (0, 3), "default": 1, "c_end": 0.5, "r_end": 0.0020}, "lmrIncCutNode": {"range": (0, 3), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "lmrIncTTCapture": {"range": (0, 3), "default": 1, "c_end": 0.5, "r_end": 0.0020}, "lmrPvNodeA": {"range": (1, 5), "default": 1, "c_end": 0.5, "r_end": 0.0020}, "lmrPvNodeB": {"range": (1, 20), "default": 11, "c_end": 1, "r_end": 0.0020}, "lmrPvNodeC": {"range": (1, 5), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "lmrCutoffCntThresh": {"range": (1, 5), "default": 3, "c_end": 0.5, "r_end": 0.0020}, "lmrIncCutoffCnt": {"range": (0, 3), "default": 1, "c_end": 0.5, "r_end": 0.0020}, "lmrStatGain": {"range": (1, 5), "default": 2, "c_end": 0.5, "r_end": 0.0020}, "lmrStatDelta": {"range": (1000, 10000), "default": 4433, "c_end": 900, "r_end": 0.0020}, "lmrRDecA": {"range": (10000, 20000), "default": 13628, "c_end": 1000, "r_end": 0.0020}, "lmrRDecB": {"range": (1000, 10000), "default": 4000, "c_end": 900, "r_end": 0.0020}, "lmrRDecDepthA": {"range": (3, 10), "default": 7, "c_end": 0.5, "r_end": 0.0020}, "lmrRDecDepthB": {"range": (11, 30), "default": 19, "c_end": 1, "r_end": 0.0020}, "lmrDeepSearchA": {"range": (10, 100), "default": 64, "c_end": 9, "r_end": 0.0020}, "lmrDeepSearchB": {"range": (1, 20), "default": 11, "c_end": 1, "r_end": 0.0020}, } PARAM_TC_INFO = { "fallingEvalA": { "range": (10, 100), "default": 66, "c_end": 10, "r_end": 0.0020, }, "fallingEvalB": { "range": (5, 25), "default": 14, "c_end": 2, "r_end": 0.0020, }, "fallingEvalC": { "range": (3, 10), "default": 6, "c_end": 1, "r_end": 0.0020, }, "fallingEvalD": { "range": (100, 1000), "default": 617, "c_end": 90, "r_end": 0.0020, }, "fallingEvalClampMin": { "range": (1, 99), "default": 51, "c_end": 10, "r_end": 0.0020, }, "fallingEvalClampMax": { "range": (101, 200), "default": 151, "c_end": 10, "r_end": 0.0020, }, "timeReductionDepth": { "range": (4, 12), "default": 8, "c_end": 0.5, "r_end": 0.0020, }, "timeReductionA": { "range": (100, 200), "default": 156, "c_end": 10, "r_end": 0.0020, }, "timeReductionB": { "range": (30, 120), "default": 69, "c_end": 10, "r_end": 0.0020, }, "timeReductionC": { "range": (100, 180), "default": 140, "c_end": 10, "r_end": 0.0020, }, "timeReductionD": { "range": (150, 250), "default": 217, "c_end": 10, "r_end": 0.0020, }, "bestMoveInstabilityA": { "range": (100, 250), "default": 179, "c_end": 15, "r_end": 0.0020, }, "totalTimeGain": { "range": (30, 100), "default": 75, "c_end": 5, "r_end": 0.0020, }, "optExtraA": { "range": (50, 150), "default": 100, "c_end": 10, "r_end": 0.0020, }, "optExtraB": { "range": (50, 150), "default": 125, "c_end": 10, "r_end": 0.0020, }, "optExtraC": { "range": (50, 150), "default": 111, "c_end": 10, "r_end": 0.0020, }, "optConstantA": { "range": (100, 500), "default": 334, "c_end": 40, "r_end": 0.0020, }, "optConstantB": { "range": (10, 50), "default": 30, "c_end": 4, "r_end": 0.0020, }, "optConstantC": { "range": (10, 100), "default": 49, "c_end": 10, "r_end": 0.0020, }, "maxConstantA": { "range": (100, 500), "default": 340, "c_end": 40, "r_end": 0.0020, }, "maxConstantB": { "range": (100, 500), "default": 300, "c_end": 40, "r_end": 0.0020, }, "maxConstantC": { "range": (100, 500), "default": 276, "c_end": 40, "r_end": 0.0020, }, "optScaleA": { "range": (100, 200), "default": 120, "c_end": 10, "r_end": 0.0020, }, "optScaleB": { "range": (20, 40), "default": 31, "c_end": 2, "r_end": 0.0020, }, "optScaleC": { "range": (30, 60), "default": 44, "c_end": 3, "r_end": 0.0020, }, "optScaleD": { "range": (10, 50), "default": 21, "c_end": 4, "r_end": 0.0020, }, "maxScaleA": { "range": (50, 100), "default": 69, "c_end": 5, "r_end": 0.0020, }, "maxScaleB": { "range": (100, 150), "default": 122, "c_end": 5, "r_end": 0.0020, }, "maximumTimeA": { "range": (50, 150), "default": 84, "c_end": 10, "r_end": 0.0020, }, "maximumTimeB": { "range": (0, 30), "default": 10, "c_end": 3, "r_end": 0.0020, }, } def print_progress_bar(current, total, bar_size=30): ratio = current / total filled_len = int(bar_size * ratio) bar = "#" * filled_len + "-" * (bar_size - filled_len) sys.stdout.write(f"\r[{bar}] {current}/{total}") sys.stdout.flush() def approximate_elo_and_error_2sigma(wins, draws, losses): total = wins + draws + losses if total == 0: return 0.0, 0.0 score = wins + 0.5 * draws p = score / total eps = 1e-9 if p <= 0.0: p = eps elif p >= 1.0: p = 1.0 - eps elo = 400.0 * math.log10(p / (1.0 - p)) try: dEdp = 400.0 / (math.log(10) * p * (1.0 - p)) sigma_p = math.sqrt(p * (1.0 - p) / total) sigma_elo = dEdp * sigma_p except ValueError: sigma_elo = 0.0 return elo, 2.0 * sigma_elo class SPSAOptimizer: def __init__( self, param_all, iterations=50000, A=5000, gamma=0.101, alpha=0.602, concurrency=32, save_step=1000, use_adam=False, test_rounds=256, ): """ param_all: 結合済みパラメータ集合。各パラメータは dict で、キー: - range, default, c_end, r_end, origin, update update==True のパラメータのみ SPSA 更新の対象となる。 """ self.iterations = iterations self.A = A self.gamma = gamma self.alpha = alpha self.concurrency = concurrency self.save_step = save_step self.use_adam = use_adam self.test_rounds = test_rounds # 結合済みパラメータ集合 self.param_all = param_all # 各パラメータの現在値 (初期値は default) self.theta = {} # SPSA 用の定数 a0, c0 (更新対象のみ) self.a0 = {} self.c0 = {} for name, info in self.param_all.items(): self.theta[name] = info["default"] if info.get("update", True): self.c0[name] = info["c_end"] * (iterations**gamma) a_end_val = info["r_end"] * (info["c_end"] ** 2) self.a0[name] = a_end_val * ((A + iterations) ** alpha) # Adam 用の内部状態（更新対象のみ） self.m = {name: 0.0 for name, info in self.param_all.items() if info.get("update", True)} self.v = {name: 0.0 for name, info in self.param_all.items() if info.get("update", True)} self.adam_t = 0 self.adam_beta1 = 0.9 self.adam_beta2 = 0.999 self.adam_epsilon = 1e-8 self.adam_lr = 0.001 self.lock = threading.Lock() self.global_iter = 0 # ログファイル作成（YYYYMMDD_HHMMSS.log） date_str = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") self.log_filename = f"log/{date_str}.log" self.log_file = open(self.log_filename, "a", buffering=1, encoding="utf-8") def worker_thread(self, thread_id, next_save_iter): while True: with self.lock: if self.global_iter >= self.iterations: return elif self.global_iter + 1 > next_save_iter: return i = self.global_iter + 1 self.global_iter = i # 各更新対象パラメータについて c(i), a(i) を計算 c_vals = {} a_vals = {} for name, info in self.param_all.items(): if info.get("update", True): c_vals[name] = self.c0[name] / (i**self.gamma) a_vals[name] = self.a0[name] / ((self.A + i) ** self.alpha) with self.lock: current_theta = self.theta.copy() # エンジンに渡すパラメータを作成（更新対象は±の擾乱、固定はそのまま） eng1_params = {} eng2_params = {} delta_dict = {} # 更新対象のパラメータの符号を記録 for name, info in self.param_all.items(): if info.get("update", True): base_val = current_theta[name] c_val = c_vals[name] delta = 1 if random.randint(0, 1) else -1 v_plus = base_val + c_val * delta v_minus = base_val - c_val * delta lo, hi = info["range"] v_plus = max(min(v_plus, hi), lo) v_minus = max(min(v_minus, hi), lo) eng1_params[name] = v_plus eng2_params[name] = v_minus delta_dict[name] = delta else: eng1_params[name] = current_theta[name] eng2_params[name] = current_theta[name] # 2ゲーム対局を実施（score = 勝ち数 - 敗北数） score = self.run_cutechess(eng1_params, eng2_params) # SPSA勾配推定と更新（更新対象のみ） with self.lock: self.adam_t += 1 for name, info in self.param_all.items(): if info.get("update", True): c_val = c_vals[name] d = delta_dict[name] g = (score / d) / (2.0 * c_val) if not self.use_adam: self.theta[name] += a_vals[name] * g else: self.m[name] = ( self.adam_beta1 * self.m[name] + (1.0 - self.adam_beta1) * g ) self.v[name] = self.adam_beta2 * self.v[name] + ( 1.0 - self.adam_beta2 ) * (g * g) m_hat = self.m[name] / (1.0 - (self.adam_beta1**self.adam_t)) v_hat = self.v[name] / (1.0 - (self.adam_beta2**self.adam_t)) adam_update = self.adam_lr * ( m_hat / (math.sqrt(v_hat) + self.adam_epsilon) ) self.theta[name] += adam_update # クリッピング lo, hi = info["range"] if self.theta[name] < lo: self.theta[name] = lo elif self.theta[name] > hi: self.theta[name] = hi print_progress_bar(self.global_iter, self.iterations) def run_cutechess(self, var_eng1, var_eng2): """ 両エンジンに対して各パラメータを渡し、2ゲーム対局を実施する。 (各エンジンは同じバイナリを呼び、名前とオプションで区別) """ engine_bin = "./cfish_params_tc" cmd = [ "cutechess-cli", "-engine", f"cmd={engine_bin}", "proto=uci", "name=plus", ] for pname, val in var_eng1.items(): ival = int(round(val)) cmd.append(f"option.{pname}={ival}") cmd += [ "-engine", f"cmd={engine_bin}", "proto=uci", "name=minus", ] for pname, val in var_eng2.items(): ival = int(round(val)) cmd.append(f"option.{pname}={ival}") cmd += [ "-each", "tc=10", "-games", "2", "-rounds", "1", "-repeat", "-openings", "file=openings.epd", "format=epd", "order=random", "-concurrency", "1", "-resign", "movecount=8", "score=600", "-draw", "movenumber=40", "movecount=8", "score=20", ] proc = subprocess.run(cmd, capture_output=True, text=True) w, l, _ = self.parse_cutechess_output(proc.stdout) return w - l def parse_cutechess_output(self, output): w, l, d = 0, 0, 0 for line in output.split("\n"): match = re.search(r"Score of .*?: (\d+) - (\d+) - (\d+)", line) if match: w, l, d = map(int, match.groups()) return w, l, d def measure_elo(self, rounds=256, concurrency=1): """ 全パラメータの現在値をエンジンに渡し、Nゲーム対局により Elo を測定する。 """ engine_bin = "./cfish_params_tc" combined = self.theta.copy() cmd = [ "cutechess-cli", "-engine", f"cmd={engine_bin}", "proto=uci", "name=Base", "-engine", f"cmd={engine_bin}", "proto=uci", "name=Test", ] for pname, val in combined.items(): ival = int(round(val)) cmd.append(f"option.{pname}={ival}") cmd += [ "-each", "tc=10", "-games", "2", "-rounds", str(rounds), "-repeat", "-openings", "file=openings.epd", "format=epd", "order=random", "-concurrency", f"{concurrency}", "-resign", "movecount=8", "score=600", "-draw", "movenumber=40", "movecount=8", "score=20", ] proc = subprocess.run(cmd, capture_output=True, text=True) w, l, d = self.parse_cutechess_output(proc.stdout) elo, elo_2s = approximate_elo_and_error_2sigma(w, d, l) return elo, elo_2s def start_optimization(self): next_save_iter = self.save_step while True: threads = [] for tid in range(self.concurrency): t = threading.Thread(target=self.worker_thread, args=(tid, next_save_iter)) threads.append(t) t.start() for t in threads: t.join() elo, elo_2sigma = self.measure_elo( rounds=self.test_rounds, concurrency=self.concurrency ) self.log_file.write(f"Iteration {self.global_iter}: Elo={elo:.2f} ±{elo_2sigma:.2f}\n") print(f"\nIteration {self.global_iter}: Elo={elo:.2f} ±{elo_2sigma:.2f}\n") self.log_file.write("Parameters:\n") for name, val in self.theta.items(): self.log_file.write(f" {name}: {val:.3f}\n") next_save_iter += self.save_step if next_save_iter > self.iterations: break print("\nSPSA Optimization Finished.") print("Final parameters:") for name, val in self.theta.items(): print(f" {name}: {val:.3f}") self.log_file.close() if __name__ == "__main__": parser = argparse.ArgumentParser(description="SPSA Optimization for Engine Parameters") # --spsa-target により更新対象を指定 parser.add_argument( "--spsa-target", choices=["both", "info", "tc"], default="info", help="更新対象。'both'なら全パラメータ更新、'info'ならPARAM_INFOのみ、'tc'ならPARAM_TC_INFOのみ更新", ) # JSONファイルによる上書き（指定されたパラメータは固定＝SPSA更新しない） parser.add_argument("--json-info", type=str, help="PARAM_INFO 用 JSON ファイル") parser.add_argument("--json-tc", type=str, help="PARAM_TC_INFO 用 JSON ファイル") parser.add_argument("--iterations", type=int, default=100000, help="SPSA の反復回数") args = parser.parse_args() # --- 各パラメータに origin と update フラグを付与（初期は update=True） --- for key, info in PARAM_INFO.items(): info["origin"] = "info" info["update"] = True for key, info in PARAM_TC_INFO.items(): info["origin"] = "tc" info["update"] = True # --spsa-target により、更新対象を制御 if args.spsa_target == "info": for key, info in PARAM_TC_INFO.items(): info["update"] = False elif args.spsa_target == "tc": for key, info in PARAM_INFO.items(): info["update"] = False # "both" の場合はそのまま # JSONファイルによる上書き (指定された場合、対象パラメータは default を上書きし、update=False とする) if args.json_info: try: with open(args.json_info, "r") as f: json_info = json.load(f) for key, val in json_info.items(): if key in PARAM_INFO: PARAM_INFO[key]["default"] = val PARAM_INFO[key]["update"] = False else: print(f"Warning: {key} は PARAM_INFO に存在しません") except Exception as e: print(f"PARAM_INFO JSON 読み込みエラー: {e}") sys.exit(1) if args.json_tc: try: with open(args.json_tc, "r") as f: json_tc = json.load(f) for key, val in json_tc.items(): if key in PARAM_TC_INFO: PARAM_TC_INFO[key]["default"] = val PARAM_TC_INFO[key]["update"] = False else: print(f"Warning: {key} は PARAM_TC_INFO に存在しません") except Exception as e: print(f"PARAM_TC_INFO JSON 読み込みエラー: {e}") sys.exit(1) # --- PARAM_INFO と PARAM_TC_INFO を結合 --- combined_params = {} for key, info in PARAM_INFO.items(): combined_params[key] = info.copy() for key, info in PARAM_TC_INFO.items(): if key in combined_params: print(f"Warning: 重複パラメータキー {key} が存在します。") combined_params[key] = info.copy() iterations = args.iterations optimizer = SPSAOptimizer( param_all=combined_params, iterations=iterations, A=(iterations * 0.1), gamma=0.101, alpha=0.602, concurrency=32, save_step=2000, use_adam=False, test_rounds=1024, ) optimizer.start_optimization()