A prototype for incremental learning finite factored planning domains from continuous perceptions and some case studies

This master project takes place within the research activity carried on by the Fondazione Bruno Kessler (FBK) in Trento, in Data and Knowledge Management unit (DKM), with Luciano Serafini as internal tutor, and Paolo Traverso as external tutor. The project focuses on automated learning of planning domains. Automated learning and planning often leverages on an abstract representation of the world called planning domains. A planning domain is described by a set of states that correspond to possible configurations of the environment, a set of actions, and a state-transition function between states, which describes the effects of actions on the environment. Planning domain models are used by agents to develop their strategies on how to act in a given environment in order to achieve their goals. The construction of these models is normally entrusted to the engineer who manually programs the agent. This work, however, requires human intervention for each new environment, while it would be desirable for an autonomous agent to be able to build a model of the environment autonomously, even if it is in an unknown environment. The aim of the research carried out by the DKM unit is to use automatic methods to learn these models by carrying out actions, and observing their effects on the environment. The specific objectives of this master thesis are to: (1) acquire specific knowledge in the field of automatic learning and planning; (2) acquire specific knowledge in the context of learning planning models; (3) extend the method developed in the research activity in FBK on states variables that are described by a set of variables taking values over domains; (4) deal with the problem of learning entirely the state-transition function; (5) extend the ALP algorithm with a prediction part to predict the effects of the actions; (6) complete the state-transition function; (7) implement the algorithm with all the new parts, provide examples, and test this extension in some simulators previously created; (8) validate and analyse the results, and measure the goodness of the model.