Multi-agent reinforcement learning for controlling a simulated compresed air system

Algoritmi umetne inteligence so v zadnjih letih močno napredovali s sposobnostjo reševanja kompleksnih problemov. Prve aplikacije te tehnologije so začele nastajati v svetu računalništva in teorije iger. Z razvojem tehnologije pa se je uporaba hitro razširila tudi na druga področja - med drugim na področje robotike in krmilnih sistemov. Vprašanje je, ali se lahko z novim pristopom umetne inteligence približamo uporabnosti konvencionalnih metod krmiljenja obsežnih sistemov za stisnjen zrak. V zaključni nalogi je predstavljena izdelava simulacijskega modela, testiranje simulacijskega modela s primerjavo z realnim sistemom, integracija spodbujevalnega učenja s simulacijskim modelom in obravnavo parametrov spodbujevalnega učenja v sistemu. Z učenjem pridobljena strategija je bila primerjana s konvencionalno metodo krmiljenja in ugotovljeno je bilo, da je učena strategija celo bolj uspešna.Artificial intelligence algorithms have made great strides in recent years with the ability to solve complex problems. The first applications of this technology began to emerge in the world of computer science and game theory. As technology evolved, the use has rapidly spread to other areas - including robotics and control systems. The question arises whether the new AI approach can bring us closer to usefulness of conventional methods of controlling large-scale compressed air systems. This thesis presents the production of a simulation model, testing of the simulation model with comparison to the real system, the integration of reinforcement learning with the simulation model and the analysis of the parameters of reinforcement learning in the system. The learned policy was compared with conventional method which proved the learned policy to be more successful