Method trainOnHistories

src/FidoControlSystem.cpp:109–142

}

double FidoControlSystem::trainOnHistories(std::vector<FidoControlSystem::History> selectedHistories, double allowedError, unsigned int maxIterations) {

	double totalError = DBL_MAX;
	net::Adadelta tempTrainer = net::Adadelta(0.95, allowedError, 100);
	unsigned int iter = 0;
	do {
		std::vector< std::vector<double> > input, correctOutput;
		for(History history : selectedHistories) {
			std::vector<Wire> historyControlWires = getWires(history.initialState);
			double newRewardForLastAction = getQValue(history.reward, history.initialState, history.newState, history.action, historyControlWires);

			Wire correctHistoryWire = {history.action, newRewardForLastAction};
			std::vector<Wire> newContolWires = newControlWires(correctHistoryWire, historyControlWires);

			input.push_back(history.initialState);
			correctOutput.push_back(getRawOutput(newContolWires));

			tempTrainer.train(network, {input.back()}, {correctOutput.back()});
		}

		totalError = 0;
		for(unsigned int a = 0; a < input.size(); a++) {
			std::vector<Wire> wires = getWires(input[a]);
			for(unsigned int b = 0; b < input[a].size(); b++) {
				totalError += pow(selectedHistories[a].reward - interpolator->getReward(wires, selectedHistories[a].action), 2);
			}
		}

		std::cout << "Error: " << totalError << "\n";
		iter++;
	} while (totalError > allowedError*selectedHistories.size() && iter < maxIterations);

	return totalError;
}

void FidoControlSystem::adjustExploration(double uncertainty) {
	explorationLevel = pow(uncertainty, 2) * 10000;