Method applyReinforcementToLastAction

src/FidoControlSystem.cpp:24–89

}

void FidoControlSystem::applyReinforcementToLastAction(double reward, State newState) {
	// SGD on current scenario
	std::vector<Wire> newContolWires = newControlWiresForHistory(History(lastState, newState, lastAction, reward));

	// Calculate uncertainty and adjust exploration
	double uncertainty = getError(lastState, getRawOutput(newContolWires));
	adjustExploration(uncertainty);
	lastUncertainty = uncertainty;

	// History sample and resize nn
	histories.push_back(History(lastState, newState, lastAction, reward));
	std::vector<History> selectedHistories = selectHistories();

	std::cout << "----------RESIZING-----------\n";
	if(selectedHistories.size() > 3) {
		bool didChange = false;
		net::Pruner pruner;
		net::NeuralNet originalNet;
		int iter = 0;
		while(iter < 1) {
			iter++;
			originalNet = net::NeuralNet(network);

			// Randomize
			for(net::Layer &l : network->net) {
				for(net::Neuron &n : l.neurons) {
					n.randomizeWeights();
				}
			}

			double totalCurrentError = trainOnHistories(selectedHistories, 0.001, 1);

			if(network->numberOfHiddenNeurons() > 4) {
				pruner.pruneRandomnly(network);
				double totalPrunedError = trainOnHistories(selectedHistories, 0.001, 1);
				if(totalPrunedError*1.05 < totalCurrentError) {
					didChange = true;
					continue;
				}
			}

			*network = originalNet;
			network->net[0].neurons.push_back(net::Neuron(network->numberOfInputs()));
			for(net::Neuron &n : network->net[1].neurons) {
				n.weights.push_back(1);
				n.randomizeWeights();
			}
			double totalAddedError = trainOnHistories(selectedHistories, 0.001, 1);
			if(totalAddedError*1.05 < totalCurrentError) {
				didChange = true;
				continue;
			}

			*network = originalNet;
			break;
		}

		if(didChange == false) {