Metis: Arming Fast Downward with Pruning and Incremental Computation

Background We consider planning tasks Π = 〈V ,O, s0, G,Cost〉 captured by the standard SAS formalism (Bäckström and Klein 1991; Bäckström and Nebel 1995) with operator costs, extended by conditional effects. In such a task, V is a set of finite-domain state variables, each with domain D(v). Each complete assignment to V is called a state, and S = ∏ v∈V D(v) is the state space of Π. The state s0 is the initial state of Π. We sometime refer to a single variable assignment as to fact. Furthermore, the goal G is a partial assignment to V , where a state s is a goal state, iff G ⊆ s1. The set O is a finite set of operators. Each operator o is given by a pair 〈pre, effs〉. The precondition pre(o) is a partial assignment to V that defines when the operator is applicable. The set effs(o) is a set of conditional effects e, each given by a pair 〈cond, eff〉 of partial assignments to V called conditions and effects. The condition cond(e) defines when the conditional effect triggers. For a shorter presentation, we assume that eff assigns a value to exactly one variable. An effect that assigns a value to more variables can be split into multiple effects. Effects that do not assign a value at all can be safely removed. Finally, Cost : O → N0 is a real-valued,