Search

U ovom radu opisane su i implementirane dvije vrste umjetne inteligencije na jednostavnom zadatku snalaženja u virtualnom okolišu, u obliku evakuacije u escape room stilu. Agenti moraju surađivati i skupljati te koristiti predmete kako bi uklonili prepreke sa svog puta prema cilju. Prvi pristup baziran je na strojnom učenju i koristi Unity ML-Agents paket s algoritmima za potporno učenje. Drugi pristup baziran je na automatskom planiranju i implementiran je jednostavan GOAP sustav za planiranje uz korištenje Unity navigacijske mreže za kretanje okolišem. Rezultat je interaktivna simulacija u kojoj korisnik može birati scenu i vrstu agenta te dodavati objekte kroz jednostavno grafičko sučelje. In this thesis two types of artificial intelligence are described and implemented to solve a task of overcoming a virtual environment in the form of an escape room. The agents have to cooperate and reach a goal while collecting items and using them to remove obstacles from their path. The first approach is machine learning based and uses the Unity ML-Agents package and its reinforcement learning algorithms. The second approach is based on automated planning and uses a simple implementation of the GOAP planning system along with the Unity NavMesh system for navigating the environment. The result is an interactive simulation in which the user can choose the scene and agent type as well as add objects through a simple graphical interface.

Ovaj diplomski rad bavi se razvojem programske podrške za planiranje i provođenje kibernetičkih vježbi. U radu su opisani procesi planiranja i provođenja kibernetičkih vježbi te je dan pregled funkcionalnih zahtjeva za programsku podršku za automatizaciju planiranja scenarija kibernetičkih vježbi. Također, opisani su i tehnički detalji implementacije programske podrške te je dan primjer korištenja. Za automatizaciju planiranja scenarija kibernetičke vježbe korišten je algoritam planiranja parcijalnog poretka te je u radu objašnjeno kako je prilagođen domeni planiranja kibernetičkih vježbi. This thesis is about development of a software for automated planning and execution of cyber exercises. It contains a description of cyber exercise planning and execution processes and it demonstrates functional requests for given software. Also, it contains a description of implementation technical details and a demonstration example. Cyber exercise scenario planning was automated using partial order planning algorithm so the thesis contains an explanation of how it was adapted to cyber exercise planning domain.

Povećanjem brige o okolišu i težnjom za smanjenjem troškova transporta problem usmjeravanja vozila (VRP) postaje sve važnija stavka u razvijenim društvima. Navedeni problem je kombinacija nekoliko klasičnih optimizacijskih problema (problem trgovačkog putnika, problem pakiranja). U radu je istraženo nekoliko inovativnih metoda koje bi se mogle primijeniti na širokom spektru problema iz stvarnog svijeta (prikupljanje otpada, dostava ...). Osnovne poteškoće na koje se nailazi su vrijeme potrebno za pronalaženje prihvatljivih rješenja i iznimno velik prostor pretraživanja rješenja. Kako bi se ostvarilo bolje rezultate od postojećih potrebno je bolje usmjeravanje prilikom pretraživanja kako se ne bi trošilo vrijeme na istraživanje nekvalitetnih rješenja. Koristeći genetičko programiranje i pohlepne funkcije moguće je brzo stvoriti početno rješenje cjelobrojnog problema usmjeravanje vozila s ciljem posluživanja određenih lokacija određenim skupom vozila, te brzo poboljšanje tako dobivenih početnih rješenja. Naknadno poboljšanje početnih rješenja moguće je opisanim paralelnim algoritmima za usmjeravanje vozila. Nakon što su dobiveni rezultati za problem usmjeravanja vozila, uočeno je da te iste rezultate ponekad nije moguće primijeniti u stvarnom svijetu. Novonastali problem riješen je stvaranjem jedinstvenog inteligentnog autonomnog prometnog sustava koji ima mogućnosti pratiti stanje prometa, otkriti moguće probleme, promijeniti stanje prometa korištenjem automatiziranog planiranja u cilju ostvarivanja bolje protočnosti prometa. Korištenjem predloženog sustava pokazano je efikasnije upravljanje prometnim sustavima. Increasingly complex variants of the vehicle routing problem with time windows (VRPTW) are coming into focus, alleviated with advances in the computing power. VRPTW is a combination of the classical travelling salesman and bin packing problems, with many real world applications in various fields – from physical resource manipulation planning to virtual resource management in the ever more popular cloud computing domain. The basis for many VRPTW approaches is a heuristic which builds a candidate solution that is subsequently improved by a search or optimization procedure. The choice of the appropriate heuristic may have a great impact on the quality of the obtained results. In this work genetic programming is used to evolve a suitable heuristic to build initial solutions for different objectives and classes of VRPTW instances. Additionally 2-phase parallel algorithm has been proposed to improve initial results obtained by genetic programming. Proposed solution is based on the divide and conquer paradigm, decomposing problem instances into smaller, mutually independent sub-problems which can be solved using traditional algorithms and integrated into a global solution of reasonably good quality. The results show great potential, since this method is applicable to different problem classes and user-defined performance objectives. It has been noticed that sometimes results for vehicle routing problem could not be used in real world applications, due to dynamic behaviour of transport systems (incidents or traffic congestion). Improving traffic control has been studied in this work. Solving traffic congestions represents a high priority issue in many big cities. Traditional traffic control systems are mainly based on pre-programmed, reactive and local techniques. This work presents an autonomic system that uses automated planning techniques instead. These techniques are easily configurable and modified, and can reason about the future implications of actions that change the default traffic lights behaviour. The proposed implemented system includes some autonomic properties, since it monitors the current traffic state, detects whether the system is degrading its performance, sets up new sets of goals to be achieved by the planner, triggers the planner that generates plans with control actions, and executes the selected courses of actions. The obtained results in several artificial and real world data-based simulation scenarios show that the proposed system can efficiently solve traffic congestion.