| Resumo: | Daily operations management is a hard task for every airline company despite the existence of earlier planning and scheduling phases that allow the elaboration of an optimal operational plan. That complexity lays on the existence of unexpected events which happen close to the day of operation. Once they are impossible to preview during the mentioned planning and scheduling phases, their occurrence may ruin the entire operational plan. In these cases it is necessary to find, as soon as possible, a solution minimizing all the costs and delay associated. This task is named as Disruption Management. The present work comes in the sequence of the project being developed at Artificial Intelligence and Computer Science Laboratory in cooperation with the TAP Portugal airline company, where a multi-agent system which aims to help at the disruption management task was developed. This system tries to find, through the use of an automated negotiation between agents, a sub-optimal solution for the operational plan's initial problem. There are two types of agents at this negotiation: agents presenting a solution proposal for a problem and another agent responsible for evaluating these same proposals and determining which is the best solution for a specific problem. The winning solution is presented to an human operator. The purpose of this work is to present a learning process that enables the agents responsible for proposing solutions to learn, along subsequent rounds, the preferences of the agent that is evaluating them. The learning process developed is adapted to the simultaneously cooperative and competitive environment where these agents are. It is also within the scope of this work to create a method of evaluation of the winning solutions. This evaluation shall be provided to the system by the human operator. To this interaction is given the name Human-in-the-Loop and its purpose is to allow the human operator to influence the evaluation function of the solutions and, therefore, the decision of the agent performing such evaluations. One of the consequences of the developed work is the achievement of more adequate solutions to both evaluator agent's preferences and to the human operator's necessities. This improvement reflects itself in the global quality of the solutions to be applied in the real context, which minimize, as much as possible, the costs and delays inherent to the change of the initial plan. The final results of this work were validated and evaluated by members of the operational control center of TAP Portugal airline company. The experiments performed allowed for a comparison between the performance of different versions of the system. The obtained results allow to affirm that the goals of this thesis were accomplished. This system is being developed with the support of the TAP Portugal airline company, which provided the needed resources to the project development, including real data related to the operational plan and occurred disruptions.
|
|---|