Institutional Repository of Vilnius University
Daugiaagentinių skatinamojo mokymosi algoritmų palyginimas simuliacinėse aplinkose /
Abstract
dc:descriptionIn this bachelor thesis an experimental comparison of multi‑agent reinforcement learning algorithms is performed using a simulation library “Melting Pot”. Three algorithms are presented: IPPO, MAPPO and QMIX. For the comparison three metrics are chosen: average total rewards, equality metric and sustainability metric. The presented algorithms are used for training multiagent strategies and then evaluated in the environments ”Commons Harvest” and ”Clean up”. The study shows that in the ”Commons Harvest” environment, the use of an individual unmodified reward signal for training is not sufficient to maximise total reward. Therefore, a study using a modified reward signal is carried out, with the best results obtained using the QMIX algorithm. The experiments also show that the use of a joint reward signal improves the average joint reward performance of the trained strategy, but makes the trained strategy unreliable. Also, it is observed that sharing the parameters of the neural networks not only reduces the learning time, but also ensures a smaller difference in the rewards received by the agents. Finally, after evaluating algorithms in the ”Clean up” environment it is evident that centralized algorithms ‑ QMIX, MAPPO perform better in this environment.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Untulis, Džiugas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD210648178&prefLang=en_US