Planejamento não determinístico baseado em redes neurais ASNet

Abstract

Automated Planning is the Artificial Intelligence subarea that is concerned with choosing an intelligent agent’s actions to achieve a goal. A planning problem in a given domain is described through an initial situation and a goal to be achieved. One solution to a planning problem is a sequence of actions that takes the agent from the initial state to a state that meets the goal, called the plan. In classical planning it is assumed that the planning environment evolves deterministically, that is, there is no uncertainty about the effects of the agent’s actions. However, there are situations where the effects of agent actions are uncertain, we may have two situations: actions with non-deterministic effects and actions with probabilistic effects. For these types of domains the solution is called policy. State of the art in automated planning consists of algorithms: (i) based on heuristic search, Boolean satisfiability, and planning graphs for deterministic action problems; (ii) based on formal techniques such as symbolic model verification for problems with non-deterministic actions and; decision-based markovian algorithms for problems with probabilistic actions. Recently, Toyer (2017) proposed the use of artificial neural networks to construct a planner for probabilistic action problems. To this end, he defined a neural network architecture called the Action Schema Network (ASNet). Subsequently, Schäfer (2018) used this architecture and proposed algorithms to obtain plans for deterministic planning problems. This paper proposes the use of ASNets neural networks to obtain policies for planning problems in domains with non-deterministic actions.