Temporal Planning while the Clock Ticks

Abstract

One of the original motivations for domain-independent planning was to generate plans that would then be executed in the environment. However, most existing planners ignore the passage of time during planning. While this can work well when absolute time does not play a role, this approach can lead to plans failing when there are external timing constraints, such as deadlines. In this paper, we describe a new approach for time-sensitive temporal planning. Our planner is aware of the fact that plan execution will start only once planning finishes, and incorporates this information into its decision making, in order to focus the search on branches that are more likely to lead to plans that will be feasible when the planner finishes.

Situated Temporal Planning Using Deadline-aware Metareasoning

“…Following Cashmore et al (2018), we formulate situated temporal planning as propositional temporal planning with Timed Initial Literals (TIL) (Cresswell and Coddington 2003;Edelkamp and Hoffmann 2004). Such problems are specified by a tuple Π = F, A, I, T, G , where:…”

“…DDA needs the D i and M i distributions as input. To estimate these distributions, we leverage estimates easily obtained in the situated temporal planner on which we build (Cashmore et al 2018). The planner uses the temporal relaxed planning graph (TRPG) (Coles et al 2010) to estimate E[D i ], and the distance to go (also from the TRPG) to estimate remaining search time (Dionne, Thayer, and Ruml 2011), which we treat as an estimate of E[M i ].…”

“…Since these errors accumulate along the path to the goal, to estimate the distribution of distance to go from state s, we convolve O d with itself d(s) times. We then multiply this distribution by the average expansion delay divided by the expansion rate (Cashmore et al 2018), which measures how much time passes, on average, between generating a state and expanding it.…”

“…The first planner to address situated temporal planning (Cashmore et al 2018) uses temporal reasoning (Dechter, Meiri, and Pearl 1991) to prune search nodes for which it is provably too late to start execution. It also uses estimates of remaining search time (Dionne, Thayer, and Ruml 2011) together with information from a temporal relaxed planning graph (Coles et al 2010) to estimate whether a search node is likely to be timely, i.e.…”

“…Finally, our third contribution is to integrate the new metareasoning scheme as the search strategy for the situated temporal planner of Cashmore et al (2018). An empirical evaluation shows that the new approach leads to timely solutions for significantly more problems than using standard heuristic search, even with pruning late nodes and dual open lists.…”