From Pattern Databases to Plan Libraries: Utilising Memory-based Methods for Improving AI Planning Performance

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

Planning is the field of Artificial Intelligence (AI) tasked with finding a sequence of actions for achieving a goal from an initial description of the environment. In this thesis, we will look into methods of leveraging memory for improving cost-optimal deterministic planning, and leveraging previous plans and executions for agents that operate in dynamic environments.

Cost-optimal deterministic planning systems have been enhanced by using heuristics into their reasoning process, but most of them are either defined by the engineers of the systems or specialised on certain tasks. The automated construction of heuristics is a long-term aim in AI Planning, and Pattern Databases (PDBs) serve as abstraction memory-based heuristics generated prior to the search to enhance computational efforts. Recent work in the automatic generation of symbolic PDBs has established it as one of the most successful approaches for cost-optimal domain-independent planning, being the approach used by the best performing planners in the last two International Planning Competitions (IPC).

We start by proposing two new approaches for combining several patterns into better heuristics, from using Constraint Programing languages (Minizinc) for finding optimal combinations, to sub-optimal but fast approaches using bin-packing algorithms for this. In continuation, we found novel ways of creating competitive patterns, by combining a full deterministic greedy algorithm we call Partial Gamer with bin-packing, which has shown that they complement very well for different types of domains.

We then changed our focus, from theoretical planning tasks to domains for robotic agents. Two major issues of these tasks arise from the stochastic nature of the environment they operate in, and from the high cost of a failure during execution, meaning frequent replanning is required. One way to address this problem is to make use of a pre-defined plan library. Such libraries have been used in Belief-Desire-Intention (BDI) agents for storing behaviour in a computationally efficient manner, however this imposes a limit on the agent’s autonomy — it can only do what its plan library allows it to do. AI Task Planning has been integrated into BDI agent programming languages to improve autonomy by allowing new plans to be created, but with limited success. We present work that combines a plan library with task planning, with results showing that such an approach alleviates the computational burden of synthesising plans, while also observing that the larger a plan library is, and to an extent the more memory used for it, the less time will be spent generating plans.
Date of Award1 Mar 2023
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorDaniele Magazzeni (Supervisor), Andrew Coles (Supervisor), Peter McBurney (Supervisor) & Stefan Edelkamp (Supervisor)

Cite this

'