Planning with Global State Constraints and State-Dependent Action Costs

Abstract

Planning with global state constraints is an extension of classical planning in which some properties of each state are derived via a set of equations, rules or constraints. This extension enables more elegant modelling of networked physical systems such as power grids. So far, research in this setting focused on domains where action costs are constant, rather than a function of a state in which the action is applied. This limitation prevents us from accurately specifying the objective in some real-world domains, leading to generation of suboptimal plans. For example, when reconfiguring a power network, we often need to temporarily leave some users without electricity for a certain amount of time, and in such circumstances it is desirable to reduce the unsupplied load over the total time span. This preference can be expressed using statedependent action costs. We extend planning with global state constraints to include state-dependent action costs, adapt abstraction heuristics to this setting, and show improved performance on a set of problems.

Urban Traffic Control via Planning with Global State Constraints (Extended Abstract)

“…This refers to an extension of classical planning in which some properties of states are determined by a set of rules that are common to all states. This formalism is well suited for applying classical planning methods on domains that involve a network of interconnected physical systems (Ivankovic, Gordon, and Haslum 2019) controlled by discrete controllable variables (for example power networks). The crucial feature of such domains is that a single discrete action (to change controllable variables), such as opening or closing a switch somewhere, affects the entire network in a way that is dependent on the global state of the system.…”

On the Compilability and Expressive Power of State-Dependent Action Costs

“…A planning task is said to have state-dependent action costs (Geißer 2018;Ivankovic, Gordon, and Haslum 2019) if the cost of one or more of its actions depends on the state where the action is applied. State-dependent action costs occur naturally, e. g., in numeric planning or in probabilistic planning in the form of state-dependent rewards of a Markov decision process (MDP).…”