我正在开发一个使用FANN（快速人工神经网络库）的软件。在多次尝试编写自己的ANN代码失败后，我尝试编译了一个FANN示例程序，这里是C++的XOR近似程序。以下是源代码。
#include "../include/floatfann.h"#include "../include/fann_cpp.h"#include <ios>#include <iostream>#include <iomanip>using std::cout;using std::cerr;using std::endl;using std::setw;using std::left;using std::right;using std::showpos;using std::noshowpos;// Callback function that simply prints the information to coutint print_callback(FANN::neural_net &net, FANN::training_data &train, unsigned int max_epochs, unsigned int epochs_between_reports, float desired_error, unsigned int epochs, void *user_data){ cout << "Epochs " << setw(8) << epochs << ". " << "Current Error: " << left << net.get_MSE() << right << endl; return 0;}// Test function that demonstrates usage of the fann C++ wrappervoid xor_test(){ cout << endl << "XOR test started." << endl; const float learning_rate = 0.7f; const unsigned int num_layers = 3; const unsigned int num_input = 2; const unsigned int num_hidden = 3; const unsigned int num_output = 1; const float desired_error = 0.001f; const unsigned int max_iterations = 300000; const unsigned int iterations_between_reports = 10000; ////Make array for create_standard() workaround (prevent "FANN Error 11: Unable to allocate memory.") const unsigned int num_input_num_hidden_num_output__array[3] = {num_input, num_hidden, num_output}; cout << endl << "Creating network." << endl; FANN::neural_net net;// cout<<"Debug 1"<<endl; //net.create_standard(num_layers, num_input, num_hidden, num_output);//doesn't work net.create_standard_array(num_layers, num_input_num_hidden_num_output__array);//this might work -- create_standard() workaround net.set_learning_rate(learning_rate); net.set_activation_steepness_hidden(1.0); net.set_activation_steepness_output(1.0); //Sample Code, changed below net.set_activation_function_hidden(FANN::SIGMOID_SYMMETRIC_STEPWISE); net.set_activation_function_output(FANN::SIGMOID_SYMMETRIC_STEPWISE); //changed above to sigmoid //net.set_activation_function_hidden(FANN::SIGMOID); //net.set_activation_function_output(FANN::SIGMOID); // Set additional properties such as the training algorithm //net.set_training_algorithm(FANN::TRAIN_QUICKPROP); // Output network type and parameters cout << endl << "Network Type : "; switch (net.get_network_type()) { case FANN::LAYER://only connected to next layer cout << "LAYER" << endl; break; case FANN::SHORTCUT://connected to all other layers cout << "SHORTCUT" << endl; break; default: cout << "UNKNOWN" << endl; break; } net.print_parameters(); cout << endl << "Training network." << endl; FANN::training_data data; if (data.read_train_from_file("xor.data")) { // Initialize and train the network with the data net.init_weights(data); cout << "Max Epochs " << setw(8) << max_iterations << ". " << "Desired Error: " << left << desired_error << right << endl; net.set_callback(print_callback, NULL); net.train_on_data(data, max_iterations, iterations_between_reports, desired_error); cout << endl << "Testing network. (not really)" << endl; //I don't really get this code --- the funny for loop. Whatever. I'll skip it. for (unsigned int i = 0; i < data.length_train_data(); ++i) { // Run the network on the test data fann_type *calc_out = net.run(data.get_input()[i]); cout << "XOR test (" << showpos << data.get_input()[i][0] << ", " << data.get_input()[i][1] << ") -> " << *calc_out << ", should be " << data.get_output()[i][0] << ", " << "difference = " << noshowpos << fann_abs(*calc_out - data.get_output()[i][0]) << endl; } cout << endl << "Saving network." << endl; // Save the network in floating point and fixed point net.save("xor_float.net"); unsigned int decimal_point = net.save_to_fixed("xor_fixed.net"); data.save_train_to_fixed("xor_fixed.data", decimal_point); cout << endl << "XOR test completed." << endl; }}/* Startup function. Synchronizes C and C++ output, calls the test function and reports any exceptions */int main(int argc, char **argv){ try { std::ios::sync_with_stdio(); // Synchronize cout and printf output xor_test(); } catch (...) { cerr << endl << "Abnormal exception." << endl; } return 0;}这是我的输出结果。
XOR test started.Creating network.Network Type : LAYERInput layer : 2 neurons, 1 bias Hidden layer : 3 neurons, 1 biasOutput layer : 1 neuronsTotal neurons and biases : 8Total connections : 13Connection rate : 1.000Network type : FANN_NETTYPE_LAYERTraining algorithm : FANN_TRAIN_RPROPTraining error function : FANN_ERRORFUNC_TANHTraining stop function : FANN_STOPFUNC_MSEBit fail limit : 0.350Learning rate : 0.700Learning momentum : 0.000Quickprop decay : -0.000100Quickprop mu : 1.750RPROP increase factor : 1.200RPROP decrease factor : 0.500RPROP delta min : 0.000RPROP delta max : 50.000Cascade output change fraction : 0.010000Cascade candidate change fraction : 0.010000Cascade output stagnation epochs : 12Cascade candidate stagnation epochs : 12Cascade max output epochs : 150Cascade min output epochs : 50Cascade max candidate epochs : 150Cascade min candidate epochs : 50Cascade weight multiplier : 0.400Cascade candidate limit :1000.000Cascade activation functions[0] : FANN_SIGMOIDCascade activation functions[1] : FANN_SIGMOID_SYMMETRICCascade activation functions[2] : FANN_GAUSSIANCascade activation functions[3] : FANN_GAUSSIAN_SYMMETRICCascade activation functions[4] : FANN_ELLIOTCascade activation functions[5] : FANN_ELLIOT_SYMMETRICCascade activation functions[6] : FANN_SIN_SYMMETRICCascade activation functions[7] : FANN_COS_SYMMETRICCascade activation functions[8] : FANN_SINCascade activation functions[9] : FANN_COSCascade activation steepnesses[0] : 0.250Cascade activation steepnesses[1] : 0.500Cascade activation steepnesses[2] : 0.750Cascade activation steepnesses[3] : 1.000Cascade candidate groups : 2Cascade no. of candidates : 80Training network.Max Epochs 300000. Desired Error: 0.001Epochs 1. Current Error: 0.25Epochs 10000. Current Error: 0.25Epochs 20000. Current Error: 0.25Epochs 30000. Current Error: 0.25Epochs 40000. Current Error: 0.25Epochs 50000. Current Error: 0.25Epochs 60000. Current Error: 0.25Epochs 70000. Current Error: 0.25Epochs 80000. Current Error: 0.25Epochs 90000. Current Error: 0.25Epochs 100000. Current Error: 0.25Epochs 110000. Current Error: 0.25Epochs 120000. Current Error: 0.25Epochs 130000. Current Error: 0.25Epochs 140000. Current Error: 0.25Epochs 150000. Current Error: 0.25Epochs 160000. Current Error: 0.25Epochs 170000. Current Error: 0.25Epochs 180000. Current Error: 0.25Epochs 190000. Current Error: 0.25Epochs 200000. Current Error: 0.25Epochs 210000. Current Error: 0.25Epochs 220000. Current Error: 0.25Epochs 230000. Current Error: 0.25Epochs 240000. Current Error: 0.25Epochs 250000. Current Error: 0.25Epochs 260000. Current Error: 0.25Epochs 270000. Current Error: 0.25Epochs 280000. Current Error: 0.25Epochs 290000. Current Error: 0.25Epochs 300000. Current Error: 0.25Testing network. (not really)XOR test (+0, -1.875) -> +0, should be +0, difference = -0XOR test (+0, -1.875) -> +0, should be +0, difference = -0XOR test (+0, +1.875) -> +0, should be +0, difference = -0XOR test (+0, +1.875) -> +0, should be +0, difference = -0Saving network.XOR test completed.训练数据（xor.data）如下所示：
4 2 1 -1 -1 -1 -1 1 1 1 -1 11 1 -1为什么ANN表现出如此诡异的学习不足现象？我非常确信在某个地方配置得非常不对劲，尤其是考虑到这是示例程序。ANN专家们，有什么建议吗？
回答：
应用FANN补丁，并确保所有对floatfann、doublefann等的引用是一致的。
