In this paper we present a new approach to the problem of planning with incomplete information and sensing. Our approach is based on a higher level, "knowledge-based," representation of the planner's knowledge and of the domain actions. In particular, in our approach we use a set of formulae from a first-order modal logic of knowledge to represent the planner's incomplete knowledge state. Actions are then represented as updates to this collection of formulae. Hence, actions are being modelled in terms of how they modify the knowledge state of the planner rather than in terms of how they modify the physical world. We have constructed a planner to utilize this representation and we use it to show that on many common problems this more abstract representation is perfectly adequate for solving the planning problem, and that in fact it scales better and supports features that make it applicable to much richer domains and problems.

Recently tremendous advances have been made in the performance of AI planning systems. However increased performance is only one of the prerequisites for bringing planning into the realm of real applications; advances in the scope of problems that can be represented and solved must also be made. In this paper we address two important representational features, concurrently executable actions with varying durations, and metric quantities like resources, both essential for modeling real applications. We show how the forward chaining approach to planning can be extended to allow it to solve planning problems with these two features. Forward chaining using heuristics or domain specific information to guide search has shown itself to be a very promising approach to planning, and it is sensible to try to build on this success. In our experiments we utilize the TLPLAN approach to planning, in which declaratively represented control knowledge is used to guide search. We show that this extra knowledge can be intuitive and easy to obtain, and that with it impressive planning performance can be achieved. 1

Partially ordered plan structures are highly suitable for centralized multi-agent planning, where plans should be minimally constrained in terms of precedence between actions performed by different agents. In many cases, however, any given agent will perform its own actions in strict sequence. We take advantage of this fact to develop a hybrid of temporal partial order planning and forward-chaining planning. A sequence of actions is constructed for each agent and linked to other agents ’ actions by a partially ordered precedence relation as required. When agents are not too tightly coupled, this structure enables the generation of partial but strong information about the state at the end of each agent’s action sequence. Such state information can be effectively exploited during search. A prototype planner within this framework has been implemented, using precondition control formulas to guide the search process. 1