""" Emergence Simulation (simplified & honest) ========================================== 問題: 前の量子版では Delta_cf がほぼ定数になった → その環境では g1 の値が適応度に影響しない 解決: Delta_cf が真に変動する環境を設計する - 信号/ノイズ比が時系列で変動 - Delta_cf_proxy = その変動の観測可能な指標 - gain = g0 + g1 * delta_cf_proxy で進化させる 示すこと: Delta_cf が変動する環境では、g1 > 0 が自然選択される (= observer-agent が evolutionary attractor) 正直な制限: これは量子模型の近似。 完全な量子版には richer operator set が必要。 """ import numpy as np import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt import matplotlib.gridspec as gridspec def make_env(n_steps=100, seed=0): """ 時系列: 真の状態 s_true と可観測ノイズ比を生成 - s_true: ゆっくり変化する4次元確率ベクトル - noise_level: 変動 (これが Delta_cf の代理指標) """ rng = np.random.default_rng(seed) s_true = np.zeros((n_steps, 4)) s_true[0] = np.array([1,0,0,0], dtype=float) for t in range(1, n_steps): # 環境がゆっくり変化 if rng.random() < 0.08: # 8%の確率でジャンプ new_s = rng.dirichlet(np.ones(4)) else: new_s = s_true[t-1] + rng.normal(0, 0.05, 4) new_s = np.abs(new_s); new_s /= new_s.sum() s_true[t] = new_s # Delta_cf proxy: 介入で未来がどれだけ分岐するかの代理指標 # signal phase (high Delta_cf): 行動が重要 # noise phase (low Delta_cf): ノイズ支配 dcf = np.zeros(n_steps) for t in range(n_steps): phase = 'signal' if (t % 30) < 18 else 'noise' if phase == 'signal': # 高Delta_cf: 介入で未来が分岐する dcf[t] = 0.20 + 0.30 * np.abs(np.sin(t * np.pi / 15)) else: # 低Delta_cf: 介入に意味がない dcf[t] = 0.02 + 0.05 * np.abs(np.sin(t * np.pi / 15)) return s_true, dcf def evaluate_individual(g0, g1, s_true, dcf, noise_amp=0.40): """ 個体 (g0, g1) の適応度を評価 gain(dcf) = clip(g0 + g1 * dcf, 0.01, 0.99) 適応度 = 真の状態に対する追跡精度（低誤差 = 高適応度） """ rng = np.random.default_rng(hash((g0, g1)) % 2**31) M = np.array([0.25, 0.25, 0.25, 0.25]) # 初期 latent state total_err = 0.0 n = len(s_true) for t in range(n): dc = dcf[t] # 観測ノイズ: Delta_cf が低い局面では観測が荒れる noise_scale = noise_amp * (1 - dc) obs = s_true[t] + rng.normal(0, noise_scale, 4) obs = np.abs(obs); obs /= obs.sum() # gain = g0 + g1 * Delta_cf g = np.clip(g0 + g1 * dc, 0.01, 0.99) # M を更新 M = (1-g) * M + g * obs M = np.abs(M); M /= M.sum() total_err += np.sum(np.abs(M - s_true[t])) return -total_err / n # 高いほど良い def evolve_population(n_steps=120, pop_size=50, n_gen=80, mutation_std=0.06, seed=42): rng = np.random.default_rng(seed) s_true, dcf = make_env(n_steps=n_steps, seed=seed+1) # 初期集団: g0 ∈ [0.1, 0.6], g1 ∈ [-0.5, 1.5] pop_g0 = rng.uniform(0.1, 0.6, pop_size) pop_g1 = rng.uniform(-0.5, 1.5, pop_size) hist_g0 = []; hist_g1 = []; hist_fit = [] for gen in range(n_gen): # 評価（複数環境シードで平均） fits = np.array([ np.mean([evaluate_individual( pop_g0[i], pop_g1[i], *make_env(n_steps, seed=seed+g*7+i%5)) for g in range(4)]) for i in range(pop_size) ]) hist_g0.append(pop_g0.mean()) hist_g1.append(pop_g1.mean()) hist_fit.append(fits.mean()) # トーナメント選択 + 変異 new_g0 = np.zeros(pop_size) new_g1 = np.zeros(pop_size) for i in range(pop_size): t = rng.choice(pop_size, size=4, replace=False) w = t[np.argmax(fits[t])] new_g0[i] = pop_g0[w] + rng.normal(0, mutation_std) new_g1[i] = pop_g1[w] + rng.normal(0, mutation_std * 1.8) pop_g0, pop_g1 = new_g0, new_g1 return hist_g0, hist_g1, hist_fit, pop_g0, pop_g1, fits, s_true, dcf def plot(hist_g0, hist_g1, hist_fit, pop_g0, pop_g1, fits, s_true, dcf): gens = np.arange(len(hist_g1)) fig = plt.figure(figsize=(14, 9)) fig.patch.set_facecolor('#0a0a1a') gs = gridspec.GridSpec(2, 2, figure=fig, hspace=0.45, wspace=0.38) def sty(ax, t): ax.set_facecolor('#0f0f2a') ax.set_title(t, color='white', pad=7, fontsize=9) ax.tick_params(colors='#888', labelsize=8) for sp in ax.spines.values(): sp.set_color('#333355') # Panel 1: g1の進化 ax=fig.add_subplot(gs[0,:]) ax.plot(gens, hist_g1, color='#44ffaa', lw=2.2, label='mean g1 (Delta_cf sensitivity)') ax.plot(gens, hist_g0, color='#5599ff', lw=1.5, alpha=0.7, label='mean g0 (base gain)') ax.axhline(0, color='white', lw=1.0, ls='--', alpha=0.6, label='g1=0 (no Delta_cf sensitivity)') g1_final = hist_g1[-1] ax.fill_between(gens,0,hist_g1, where=np.array(hist_g1)>0, color='#44ffaa',alpha=0.12) ax.annotate(f'converged g1={g1_final:.3f} > 0\n=> observer-agent naturally selected', xy=(gens[-1],g1_final), xytext=(gens[-1]*0.6, g1_final+0.2), color='#44ffaa',fontsize=8.5, arrowprops=dict(arrowstyle='->',color='#44ffaa',lw=1.2)) sty(ax,'Evolution of gain sensitivity (g1)\ng1 > 0 means: larger update when Delta_cf is high (observer-agent regime)') ax.set_xlabel('generation',color='#aaa',fontsize=8) ax.set_ylabel('g1 / g0',color='#aaa',fontsize=8) ax.legend(fontsize=8,facecolor='#111133',labelcolor='white',loc='lower right') # Panel 2: 最終世代の散布図 ax=fig.add_subplot(gs[1,0]) sc=ax.scatter(pop_g0,pop_g1,c=fits,cmap='plasma',s=35,alpha=0.8) ax.axhline(0,color='white',lw=0.8,ls='--',alpha=0.5) plt.colorbar(sc,ax=ax,label='fitness') sty(ax,'Final population distribution\n(g1 > 0 zone = observer-agent region)') ax.set_xlabel('g0 (base gain)',color='#aaa',fontsize=8) ax.set_ylabel('g1 (Delta_cf sensitivity)',color='#aaa',fontsize=8) top_fits = np.percentile(fits,75) mask = fits > top_fits ax.scatter(pop_g0[mask],pop_g1[mask],c='white',s=80,alpha=0.5, marker='*',label='top 25%') ax.legend(fontsize=7.5,facecolor='#111133',labelcolor='white') # Panel 3: Delta_cf の時系列 ax=fig.add_subplot(gs[1,1]) steps=np.arange(len(dcf)) ax.plot(steps,dcf,color='white',lw=1.4,alpha=0.9) ax.fill_between(steps,0,dcf,color='#44ffaa',alpha=0.15) ax.axhline(0.1,color='#ffaa33',lw=0.8,ls='--',alpha=0.6, label='low Delta_cf region') sty(ax,'Delta_cf (env. informativeness)\n' 'alternates between signal/noise phases') ax.set_xlabel('step',color='#aaa',fontsize=8) ax.set_ylabel('Delta_cf proxy',color='#aaa',fontsize=8) ax.legend(fontsize=7.5,facecolor='#111133',labelcolor='white') fig.suptitle( 'Selection Principle: Evolutionary emergence of Delta_cf-sensitive updating\n' 'Environment with variable informativeness selects g1 > 0 (observer-agent)', color='white',fontsize=10,y=0.99) plt.savefig('/mnt/user-data/outputs/emergence_simulation.png', dpi=150,bbox_inches='tight',facecolor=fig.get_facecolor()) plt.close() if __name__ == '__main__': print("Running emergence simulation (simplified env)...") r = evolve_population(n_steps=120, pop_size=50, n_gen=80, seed=42) hist_g0,hist_g1,hist_fit,pop_g0,pop_g1,fits,s_true,dcf = r g1f = hist_g1[-1]; g0f = hist_g0[-1] print(f"\nFinal: g0={g0f:.4f}, g1={g1f:.4f}") print() if g1f > 0.05: print(">>> g1 > 0: Delta_cf-sensitive update naturally selected") print(f" gain at dcf=0.0: {np.clip(g0f,0.01,0.99):.3f}") print(f" gain at dcf=0.3: {np.clip(g0f+g1f*0.3,0.01,0.99):.3f}") print(f" gain at dcf=0.5: {np.clip(g0f+g1f*0.5,0.01,0.99):.3f}") print("\n -> Selection principle demonstrated in simplified environment.") print(" -> Full quantum version needs richer operator set (noted limitation).") else: print(">>> g1 ~ 0: neutral in this environment") # top 10 individuals top10 = np.argsort(fits)[-10:] print(f"\nTop 10 individuals: mean g1={pop_g1[top10].mean():.4f}") plot(hist_g0,hist_g1,hist_fit,pop_g0,pop_g1,fits,s_true,dcf) print("\nDone.")