|
|
The use of reinforcement learning in abstract route selection
GLASER, Leoš
The thesis deals with reinforcement learning approach for designing a route for an agent in a simplified scenario of movement in a transportation network. The theoretical part introduces the basics of artificial intelligence, reinforcement learning, and selected methods of reinforcement learning, both classical and modern. Additionally, the basic theory related to traffic simulation is briefly mentioned. In the practical part of the thesis, a console application utilizing selected reinforcement learning methods is developed. The methods are used to design a waste collection route in a selected district in České Budějovice and compared to a method solving this task using swarm intelligence. The results of the reinforcement learning-designed routes are similar to the results obtained by swarm intelligence, with Proximal Policy Optimization with action masking being the most successful method overall. In one case, an optimal solution is found.
|
|
|
Support of management decision-making on transport networks
Přibyl, Vladimír ; Černý, Jan (advisor) ; Kavička, Antonín (referee) ; Peško, Štefan (referee)
The presented thesis is focused on a set of problems related to managerial decision-making concerning networks (particularly transportation networks), respectively - if we put it more precisely - the thesis focuses on the support of this decision-making by means of quantitative methods. A set of problems related to nets and decision-making concerning their individual parts or elements represents a very complex sphere which has been a subject of research for a number of decades. Out of this sphere, the thesis formulates and elaborates in great detail two problems, which - from the point of view of their practical significance - are important for the decision-making of managers of carriers, or the public sphere, and which have not been published in this form yet. The main point is the problem of how to find a subnet with a limited prolongation of routes between important pairs of vertices. Another problem is a design of a bus route in an area with a low demand. For each of these problems, the thesis offers an exact combinatorial solution method, furthermore a method based on integer linear programming, and - last but not least - also, of course, heuristic methods of solution. All these methods have been tested on a set of networks, which has been created for this purpose in a pseudo-random way in the frame of this thesis. The testing has been focused primarily on the comparison of the results provided by heuristic methods, which are of great importance - with regard to a great computational difficulty of exact methods - for feasible tasks on a larger scale. The tests have proved that the proposed heuristic methods are practically applicable and show results whicheven represent the optimal solution in a number of cases, or are only slightly distant from the optimal solution.
|
guest ::
