|RAMPONI GIORGIA||Cycle: XXXIII |
Section: Computer Science and Engineering
Tutor: PERNICI BARBARA
Advisor: RESTELLI MARCELLO Major Research topic
:Multi-agent Reinforcement Learning: dealing with partial informationAbstract:
A Multi-Agent System (MAS) is a set of agents that co-exist in the same environment. In MAS the agents can interact cooperating, competing or influencing each other. The applications of Multiagent systems are extremely wide and include: energy applications, robot controlling, autonomous driving problems, security, marketing, advertising. An example is to implement an electricity network or, for instance, a Social Network is a Multi-Agent system where more accounts (agents) share posts, comment and create connections (take an action in the environment). In these environments, agents take decisions considering their interest and the reactions of other agents to their actions. Reinforcement Learning is an algorithmic paradigm that aims to create artificial agents that can learn how to sequentially act to optimize their utility. In Multiagent Reinforcement Learning the environment is not stationary anymore, since the result of an agent’s action depends on other agents. This leads to the failure of the standard Reinforcement Learning algorithms that are based on the stationarity assumption. Although in last years many Multi-Agent Reinforcement Learning algorithms were proposed to address the multi-agent setting, the majority of them do not take into account the complex interactions between the agents. To incorporate in the decision-making algorithm the other agents, it is important to understand the other agents’ future actions, which change by their learning algorithms and reward functions. Then we have to use this information to construct a MARL algorithm. The status of the art in MARL considers the setting in which all of this information is known. In my research we remove this assumption answering to two main research questions: 1. How to learn the intentions of other agents? 2. How to use the learned intentions to construct learning agents?