Goal reasoning for autonomous agents using automated planning

• Autores: Alberto Pozanco Lancho
• Directores de la Tesis: Daniel Borrajo Millán (dir. tes.) , Susana Fernández Arregui (codir. tes.)
• Lectura: En la Universidad Carlos III de Madrid ( España ) en 2021
• Idioma: español
• Texto completo no disponible (Saber más ...)
• Resumen

  • Automated planning deals with the task of finding a sequence of actions, namely a plan, which achieves a goal from a given initial state. Most planning research consider goals are provided by a external user, and agents just have to find a plan to achieve them. However, there exist many real world domains where agents should not only reason about their actions but also about their goals, generating new ones or changing them according to the perceived environment. In this thesis we aim at broadening the goal reasoning capabilities of planning-based agents, both when acting in isolation and when operating in the same environment as other agents.

    In single-agent settings, we firstly explore a special type of planning tasks where we aim at discovering states that fulfill certain cost-based requirements with respect to a given set of goals. By computing these states, agents are able to solve interesting tasks such as find escape plans that move agents in to safe places, hide their true goal to a potential observer, or anticipate dynamically arriving goals. We also show how learning the environment's dynamics may help agents to solve some of these tasks. Experimental results show that these states can be quickly found in practice, making agents able to solve new planning tasks and helping them in solving some existing ones.

    In multi-agent settings, we study the automated generation of goals based on other agents' behavior. We focus on competitive scenarios, where we are interested in computing counterplans that prevent opponents from achieving their goals. We frame these tasks as counterplanning, providing theoretical properties of the counterplans that solve them. We also show how agents can benefit from computing some of the states we propose in the single-agent setting to anticipate their opponent's movements, thus increasing the odds of blocking them. Experimental results show how counterplans can be found in different environments ranging from competitive planning domains to real-time strategy games.