{ "device": "cuda", "math_check": { "rows": [ { "n": 2, "tau": 0.2, "softmin_error": 0.01654753088951111, "softmax_error": 0.01654753088951111, "bound": 0.13862943611198905, "min_bound_ratio": 0.9914180684697812, "max_bound_ratio": 0.9914189283830356 }, { "n": 2, "tau": 0.1, "softmin_error": 0.004213903099298477, "softmax_error": 0.004213901236653328, "bound": 0.06931471805599453, "min_bound_ratio": 0.982887728986986, "max_bound_ratio": 0.982887728986986 }, { "n": 2, "tau": 0.05, "softmin_error": 0.0011359817581251264, "softmax_error": 0.0011359798954799771, "bound": 0.03465735902799726, "min_bound_ratio": 0.9659818344071719, "max_bound_ratio": 0.9659818344071719 }, { "n": 2, "tau": 0.02, "softmin_error": 0.0002304219815414399, "softmax_error": 0.00023042074462864548, "bound": 0.013862943611198907, "min_bound_ratio": 0.9164955464479072, "max_bound_ratio": 0.9164869473153641 }, { "n": 4, "tau": 0.2, "softmin_error": 0.03395505249500275, "softmax_error": 0.035028714686632156, "bound": 0.2772588722239781, "min_bound_ratio": 0.79778866002239, "max_bound_ratio": 0.7993829929404667 }, { "n": 4, "tau": 0.1, "softmin_error": 0.008019108325242996, "softmax_error": 0.008443870581686497, "bound": 0.13862943611198905, "min_bound_ratio": 0.6464604268486855, "max_bound_ratio": 0.6500396007914542 }, { "n": 4, "tau": 0.05, "softmin_error": 0.0018046360928565264, "softmax_error": 0.0019981705117970705, "bound": 0.06931471805599453, "min_bound_ratio": 0.4714345429788595, "max_bound_ratio": 0.49639030051062094 }, { "n": 4, "tau": 0.02, "softmin_error": 0.0002286528324475512, "softmax_error": 0.0002826704876497388, "bound": 0.027725887222397813, "min_bound_ratio": 0.3626383180430467, "max_bound_ratio": 0.4893215448372675 }, { "n": 8, "tau": 0.2, "softmin_error": 0.049777381122112274, "softmax_error": 0.04956043139100075, "bound": 0.4158883083359672, "min_bound_ratio": 0.5560633358588032, "max_bound_ratio": 0.5959898248509828 }, { "n": 8, "tau": 0.1, "softmin_error": 0.011563748121261597, "softmax_error": 0.011731680482625961, "bound": 0.2079441541679836, "min_bound_ratio": 0.4159583859277947, "max_bound_ratio": 0.551861656379349 }, { "n": 8, "tau": 0.05, "softmin_error": 0.0025859149172902107, "softmax_error": 0.002830208744853735, "bound": 0.1039720770839918, "min_bound_ratio": 0.3062644115544781, "max_bound_ratio": 0.524171876314934 }, { "n": 8, "tau": 0.02, "softmin_error": 0.0003214461321476847, "softmax_error": 0.0004623171698767692, "bound": 0.04158883083359671, "min_bound_ratio": 0.2655469456872053, "max_bound_ratio": 0.5053918007973425 }, { "n": 16, "tau": 0.2, "softmin_error": 0.06473999470472336, "softmax_error": 0.06601116806268692, "bound": 0.5545177444479562, "min_bound_ratio": 0.5083207370009739, "max_bound_ratio": 0.5319373946174548 }, { "n": 16, "tau": 0.1, "softmin_error": 0.015401076525449753, "softmax_error": 0.015827631577849388, "bound": 0.2772588722239781, "min_bound_ratio": 0.44194424787878656, "max_bound_ratio": 0.4317413771163504 }, { "n": 16, "tau": 0.05, "softmin_error": 0.003675160463899374, "softmax_error": 0.0037773565854877234, "bound": 0.13862943611198905, "min_bound_ratio": 0.3548457841324529, "max_bound_ratio": 0.3078764622685655 }, { "n": 16, "tau": 0.02, "softmin_error": 0.0005827514105476439, "softmax_error": 0.000571474083699286, "bound": 0.055451774444795626, "min_bound_ratio": 0.24365212104361256, "max_bound_ratio": 0.23940629935043697 }, { "n": 32, "tau": 0.2, "softmin_error": 0.07556750625371933, "softmax_error": 0.07577236741781235, "bound": 0.6931471805599454, "min_bound_ratio": 0.42774493026692584, "max_bound_ratio": 0.4306696672274987 }, { "n": 32, "tau": 0.1, "softmin_error": 0.016352545469999313, "softmax_error": 0.01680590771138668, "bound": 0.3465735902799727, "min_bound_ratio": 0.3244786354869558, "max_bound_ratio": 0.335832242166315 }, { "n": 32, "tau": 0.05, "softmin_error": 0.003503528656437993, "softmax_error": 0.003680023830384016, "bound": 0.17328679513998635, "min_bound_ratio": 0.2686479075373623, "max_bound_ratio": 0.248648389033828 }, { "n": 32, "tau": 0.02, "softmin_error": 0.0004660064005292952, "softmax_error": 0.00044423865620046854, "bound": 0.06931471805599453, "min_bound_ratio": 0.1872581499124433, "max_bound_ratio": 0.17768903521843663 } ], "max_bound_ratio_min": 0.9914180684697812, "max_bound_ratio_max": 0.9914189283830356, "fixed_tau": [ 0.4000000059604645, 0.20000000298023224, 0.10000000149011612, 0.05000000074505806, 0.02500000037252903 ], "fixed_softmin_abs_error": [ 0.18702058494091034, 0.042145296931266785, 0.004870489239692688, 0.00012378394603729248, 1.4901161193847656e-07 ] }, "results": [ { "seed": 11, "beta": 0.0, "prediction_mse": 0.0005870691384188831, "hard_max_residual": 0.07603210210800171, "hard_mean_residual": -0.0004309859359636903, "coverage_at_eps": 0.98046875, "avg_successors_at_eps": 1.296875, "zero_coverage_fraction": 0.00390625, "smooth_sim": 0.31886938214302063 }, { "seed": 22, "beta": 0.0, "prediction_mse": 0.0007153446786105633, "hard_max_residual": 0.07124549150466919, "hard_mean_residual": 0.00039630423998460174, "coverage_at_eps": 1.0, "avg_successors_at_eps": 1.3515625, "zero_coverage_fraction": 0.0, "smooth_sim": 0.3169763386249542 }, { "seed": 33, "beta": 0.0, "prediction_mse": 0.0006021782755851746, "hard_max_residual": 0.06473100185394287, "hard_mean_residual": 0.00022710356279276311, "coverage_at_eps": 1.0, "avg_successors_at_eps": 1.34375, "zero_coverage_fraction": 0.0, "smooth_sim": 0.31762558221817017 }, { "seed": 11, "beta": 0.15, "prediction_mse": 0.002313392236828804, "hard_max_residual": 0.03756055235862732, "hard_mean_residual": -0.11730918288230896, "coverage_at_eps": 0.1953125, "avg_successors_at_eps": 0.8828125, "zero_coverage_fraction": 0.31640625, "smooth_sim": 0.23769232630729675 }, { "seed": 22, "beta": 0.15, "prediction_mse": 0.0015411432832479477, "hard_max_residual": 0.0019683241844177246, "hard_mean_residual": -0.18119104206562042, "coverage_at_eps": 0.07421875, "avg_successors_at_eps": 0.45703125, "zero_coverage_fraction": 0.54296875, "smooth_sim": 0.19591264426708221 }, { "seed": 33, "beta": 0.15, "prediction_mse": 0.0039921416901052, "hard_max_residual": -0.14899271726608276, "hard_mean_residual": -0.23391635715961456, "coverage_at_eps": 0.0, "avg_successors_at_eps": 0.609375, "zero_coverage_fraction": 0.39453125, "smooth_sim": 0.10364626348018646 }, { "seed": 11, "beta": 0.5, "prediction_mse": 0.0153421089053154, "hard_max_residual": 0.004907041788101196, "hard_mean_residual": -0.16108031570911407, "coverage_at_eps": 0.1171875, "avg_successors_at_eps": 1.328125, "zero_coverage_fraction": 0.17578125, "smooth_sim": 0.17853860557079315 }, { "seed": 22, "beta": 0.5, "prediction_mse": 0.03535536676645279, "hard_max_residual": -0.20509329438209534, "hard_mean_residual": -0.3234386444091797, "coverage_at_eps": 0.0, "avg_successors_at_eps": 0.7265625, "zero_coverage_fraction": 0.33203125, "smooth_sim": 0.023281831294298172 }, { "seed": 33, "beta": 0.5, "prediction_mse": 0.028807854279875755, "hard_max_residual": -0.245570570230484, "hard_mean_residual": -0.3344379663467407, "coverage_at_eps": 0.0, "avg_successors_at_eps": 0.69921875, "zero_coverage_fraction": 0.30078125, "smooth_sim": 0.003341417293995619 } ] }