Upload results/benchmark_200K_a10g_2026-05-05.json

results/benchmark_200K_a10g_2026-05-05.json

@@ -0,0 +1,87 @@
+{
+  "config": {
+    "max_rows": 200000,
+    "budget": 10000.0,
+    "T": 10000,
+    "vpc": 50.0,
+    "k": 0.8,
+    "n_runs": 5,
+    "seed": 42,
+    "ctr_test_auc": 0.6947,
+    "algorithms": ["DualOGD", "TwoSidedDual", "ValueShading", "RLB", "Linear", "Threshold"],
+    "run_date": "2026-05-05",
+    "hardware": "a10g-small",
+    "data_size": "200K rows (Criteo_x4)",
+    "auction_type": "first-price",
+    "market_price": "log-normal conditioned on pCTR features"
+  },
+  "aggregated": {
+    "TwoSidedDual": {
+      "clicks_mean": 284.6,
+      "clicks_std": 8.3,
+      "cpc_mean": 33.41,
+      "cpc_std": 0.86,
+      "budget_used_mean": 0.950,
+      "budget_used_std": 0.005,
+      "win_rate_mean": 0.076,
+      "win_rate_std": 0.002
+    },
+    "ValueShading": {
+      "clicks_mean": 257.8,
+      "clicks_std": 7.4,
+      "cpc_mean": 38.82,
+      "cpc_std": 1.14,
+      "budget_used_mean": 1.0,
+      "budget_used_std": 0.0,
+      "win_rate_mean": 0.082,
+      "win_rate_std": 0.002
+    },
+    "DualOGD": {
+      "clicks_mean": 248.0,
+      "clicks_std": 9.4,
+      "cpc_mean": 31.18,
+      "cpc_std": 1.13,
+      "budget_used_mean": 0.773,
+      "budget_used_std": 0.027,
+      "win_rate_mean": 0.066,
+      "win_rate_std": 0.002
+    },
+    "RLB": {
+      "clicks_mean": 135.8,
+      "clicks_std": 13.3,
+      "cpc_mean": 74.34,
+      "cpc_std": 7.16,
+      "budget_used_mean": 1.0,
+      "budget_used_std": 0.0,
+      "win_rate_mean": 0.042,
+      "win_rate_std": 0.004
+    },
+    "Threshold": {
+      "clicks_mean": 71.0,
+      "clicks_std": 4.1,
+      "cpc_mean": 70.36,
+      "cpc_std": 3.96,
+      "budget_used_mean": 0.0,
+      "budget_used_std": 0.0,
+      "win_rate_mean": 0.017,
+      "win_rate_std": 0.001
+    },
+    "Linear": {
+      "clicks_mean": 63.6,
+      "clicks_std": 6.0,
+      "cpc_mean": 79.20,
+      "cpc_std": 6.17,
+      "budget_used_mean": 0.0,
+      "budget_used_std": 0.0,
+      "win_rate_mean": 0.020,
+      "win_rate_std": 0.003
+    }
+  },
+  "notes": {
+    "data_note": "Linear and Threshold show 0% budget used because they were configured with infinite internal budget. They actually spent ~5000 each. The set_budget call is being fixed in a follow-up.",
+    "ctr_model": "LogisticRegression (max_iter=500, C=0.1), AUC=0.6947",
+    "two_sided_dual_advantage": "The TwoSidedDual beats DualOGD by spending 95% vs 77% of budget, converting 15% more clicks. The floor multiplier ν prevents the dual from becoming too conservative.",
+    "rl_performance": "RLB underperforms because it needs more data to learn its Q-table. With only 10K auctions per run, the tabular approach hasn't covered the state space.",
+    "citation": "All algorithms documented in RESEARCH_RESOURCES.md. Primary paper: Wang et al. 2023, arXiv:2304.13477"
+  }
+}