Application of neural networks in finding a game solution

Unity 3D, jednostavan program za izradu igara, uz ML-Agents dodatak pruža nove usluge treniranja umjetne inteligencije da izvršava određene zadatke. Treniranje agenata ne ovisi samo o jačini CPU ili GPU, već ovisi i o okruženju koje stvorimo za agente te mehanici nagrađivanja koju napišemo u kodu. Agenti uče uz pomoć neuronske mreže kojoj mi određujemo broj ulaza, odnosno opservacija te broj izlaza, odnosno akcija koje agent izvršava. Vrlo je bitno agenta postepeno učiti od lakših pa sve do težih zadataka. Da bismo trenirali agente kroz Unity i ML-Agents dodatak, nije potrebno dodatno znanje o neuronskim mrežama zbog toga što nam funkcije neuronske mreže izvršava TensorFlow u pozadini, algoritmom PPO kojeg mi ne mijenjamo.Unity 3D, simple program for creating games, with ML-Agents addition gives new services of training artificial intelligence to do the specific tasks. Training agents do not depend only on strength of CPU or GPU, but it depends more on the environment which we create for agents and also the mechanic of rewarding that we write in code. Agents learn throughout the help of neuron network whom we determine the number of inputs, also called observations and the number of outputs, also called actions that the agent needs to do. It is very important for the agent to be learnt step by step from very easy to very hard tasks. To train agents throughout Unity and ML-Agents, it is not necessary to know additional knowledge about neural networks because all functions of the neural network are done by TensorFlow in the background, by algorithm PPO which we do not change by hand