Behavior training of physical agents using neural networks

U ovom radu je prikazan program u kojem se nalazi neuronska mreža i kod za upravljanje robotom. Uz pomoć algoritma povratne propagacije konstruiramo neuronsku mrežu. Pokušavamo eksperimentalnim putem odrediti optimalni broj neurona u skrivenom sloju i broj potrebnih epoha, jer ne postoji jedinstvena formula za određivanje tih parametara. Ako je broj epoha prevelik, doći će do pretreniranosti mreže, a ako je premalen, mreža se neće uspjeti istrenirati. Ulazni set podataka za treniranje mreže je osmišljen da robot koji simbolizira psa ima logične izlaze s obzirom na ulaze, tako na primjer, ako pas vidi crvenu boju i rep mu se dira akcija psa će biti lajanje. Kad smo mrežu istrenirali, primjenjujemo je na psu koji predstavlja fizički dio agenta. Ulazi će se očitavati sa senzora, a akcije dobivamo pomoću motora i zvuka koji se nalaze na robotu.This thesis describes the creation and implementation of a program which includes a neural network and the code to control a physical agent – a robot dog. A neural network is created by using the backpropagation algorithm. Since a unique mathematical expression to determine the optimal number of hidden layer neurons and the required number of epochs doesn't exist, the experimental part of the paper focuses on determining those parameters. If the number of epochs is too large, the network will become overtrained and in the case of selecting a too small number of epochs, the network will be unable to train itself. The input set of data required to train the network is designed in such a way that a robot, representing a dog, has logical outputs in regard to the inputs. For example, if the dog percieves the color red and its tail is touched, it will respond with barking. Once the training is complete, the trained network is applied to the dog, which represents the physical part of the agent. The input data is fed through the dog's sensors and the actions are performed by it's built in motors and sounds