Augmenting BDI Agents with
Deliberative Planning Techniques
Andrzej Walczak, Lars Braubach, Alexander Pokahr, and Winfried Lamersdorf
Distributed Systems and Information Systems
Department of Informatics, University of Hamburg
D-22527 Hamburg, Germany
Abstract. Belief-Desire-Intention (BDI) agents are well suited for au-
tonomous applications in dynamic environments. Their precompiled plan
schemata contain the procedural knowledge of an agent and contribute to
the performance. The agents generally are constrained to a fixed set of ac-
tion patterns. Their choice depends on current goals, not on the future of
the environment. Planning techniques can provide dynamic plans regard-
ing the predicted state of the environment. We augment a BDI frame-
work with a state-based planner for operational planning in domains
where BDI is not well applicable. For this purpose, the requirements for
the planner and for the coupling with a BDI system are investigated. An
approach is introduced where a BDI system takes responsibility for plan
monitoring and re-planning and the planner for the creation of plans. A
fast state-based planner utilizing domain specific control knowledge re-
tains the responsiveness of the system. In order to facilitate integration
with BDI systems programmed in object-oriented languages, the plan-
ning problem is adopted into the BDI conceptual space with object-based
domain models. The application of the hybrid system is illustrated using
a propositional puzzle and a multi agent coordination scenario.
1 Introduction
BDI is a well established model of agency [1] based on the Theory of Practical
Reasoning [2]. Early BDI-systems have been devised with the idea in mind to
overcome the poor performance of propositional planners controlling agents in
dynamic environments at that time. The systems are based on two central ideas.
One of them is the reactive planning, comparable with hierarchical planning
systems [3], the other is deliberation [4,5].
Planning, an approach central to Artificial Intelligence (AI) research, is sub-
stantial for rational agent behavior. It is a method that aids agents in solving
complex problems in synthetic and natural environments. Although planning
systems are devised for means-end reasoning and are capable to find actions
that achieve goals, they are less useful to decide, which goals to pursue [6].
Due to advances in planning techniques and understanding of planning prob-
lems, it seams reasonable and interesting to combine the strength of flexible
means-end reasoning given by deliberative planners with the timely reactivity
R.H. Bordini et al. (Eds.): ProMAS 2006, LNAI 4411, pp. 113–127, 2007.
c© Springer-Verlag Berlin Heidelberg 2007

114 A. Walczak et al.
and goal deliberation capabilities carried by BDI systems. It is also interest-
ing to analyze suitability of planning approaches to BDI agents in real world
applications.
In order to benefit from both paradigms one needs to consider the strengths
of both paradigms. There are multiple ways to compose the systems and the
outcome is different dependent on their properties and features. As stated above,
hierarchical planning techniques are comparable with BDI. Their strength lies
in the evaluation of future environmental states and a constructed proof that a
course of action will achieve goals under the preconditions. On the other hand,
BDI systems handle dynamic environments more efficiently and are capable of
both: reactive behavior, and maintenance of long term goals. They sacrifice the
optimality and correctness of their planning algorithms for performance.
Both paradigms deal with generation of actions andsharecommonideas,so
there are concerns which parts of a control and planning problem will be dele-
gated to a planner and which to the BDI subsystem. This determines the choice
of the BDI component and the planning algorithm. The overall architecture
depends strongly on those choices.
A hybrid system can be built twofold. The planner may be applied to produce
long term plans and to hand over single parts to a reactive BDI subsystem for
the execution. This approach invokes serious performance concerns especially in
dynamic environments where continuous changes force the planner to re-plan -
a process with performance penalties comparable to planning itself. Generally,
planning algorithms have been devised for one shot planning and are well suited
for a solution of a single planning problem. They are rather less useful to maintain
long term intentions of an agent.
The other way round is to augment the BDI system with a relatively simple
planner that is invoked from the BDI controller and used for the purpose of
creating short-term plans that need a proof of correctness. The last approach is
used in this work to join the best from both paradigms.
The remainder of this paper proceeds as follows. In Section 2 we define the
concept of a planning problem used for this work. Section 3 discusses the choice
of a planning algorithm. In Section 4 we propose a way to integrate a planning
component into a BDI framework. Section 5 presents two application examples
of the hybrid system. Related work is presented in Section 6 and a conclusion is
giveninSection7.
2 Planning Concepts
The basis for planning is given in the form of a planning problem.Inorderto
represent a planning problem one needs at least to describe states of the world
and how these states may change due to agent’s actions. In a restricted classical
view, this can be given by a model of a state-transition system Σ =(S,A,γ)
where S is the set of states, A is the set of actions and γ : S × A → S ∪{⊥}
is the transition function mapping a state and action to another state. ⊥ is the
illegal state being a result of a not applicable action. The planning problem is