Norm-Aware Planning: Semantics and Implementation Sofia Panagiotidi Universitat Polit` ecnica de Catalunya C/ Jordi Girona 1-3, 08034 Barcelona (Spain) Email: panagiotidi@lsi.upc.edu Javier V´ azquez-Salceda Universitat Polit` ecnica de Catalunya C/ Jordi Girona 1-3, 08034 Barcelona (Spain) Email: jvazquez@lsi.upc.edu Abstract—Norms are a way to specify acceptable behaviour in a context. In literature there is a lot of work on norm theories, models and specifications on how agents might take norms into account when reasoning but few practical implementations. In this paper we present a framework and an implementation for norm-oriented planning. Unlike most frameworks, our approach takes into consideration the operationalisation of norms during the plan generation phase. In our framework norms can be obligations or prohibitions which can be violated, and are accompanied by repair norms in case they are breached. Norm operational semantics is expressed as an extension/on top of STRIPS semantics, acting as a form of temporal restrictions over the trajectories (plans) computed by the planner. In combination with the agent’s utility functions over the actions, the norm-aware planner computes the most profitable trajectory concluding to a state of the world where no pending obligations exist and any (obligation/prohibition) violation has been handled. An implementation of the framework in PDDL is provided. I. I NTRODUCTION -BACKGROUND In the last decade, computational systems have become more and more complex resulting in highly complicated in- terconnected networks. As systems grow to include hundreds of components, the need to design organisational autonomous models where the members’ behaviour is somehow regulated in order to produce desirable and avoid undesirable situations is becoming stronger. In such models, goals and interactions are usually not specified in terms of the mental states of individual agents, but in terms of organisational concepts such as roles (or function, or position), groups (or communities), norms (or regulations or policies) and communication proto- cols (including ontologies). In these cases, agents are seen as actors that perform the role(s) described by the organisation design. Often the very notion of agent autonomy refers to the capability of an agent to act independently, exhibiting control over its own internal state meaning that an agent needs to anticipate, plan and adopt actions that are in accordance to organisational specifications while at the same time optimising its individual outcome. Despite the fact that such norm auton- omy is important in complex systems where dynamic decision making is a key element and where conflicts frequently occur, few systems achieve capturing an efficient normative reasoning procedure. Envisioning such a functionality, this paper describes in detail a framework to support practical normative reasoning that can be used by agents to produce their plans. This is done by extending the well known STRIPS language to include an extra layer of normative representation that, on top of the domain description, adds norms that act as complex restrictions to the planning problem and influence the planning mechanism in a way that it produces the most beneficial plans. In this way, a standard PDDL planner is able to receive the domain description and additionally the normative layer and to compute executional paths that consider restrictions imposed by norms, either conforming to or avoiding to take into consideration while considering any possible sanction. Similar to our work is PDDL 3.0 [1], an extension of planning language PDDL (originally implementing STRIPS) that imposes strong and soft constraints expressed in Linear Temporal Logic formulas on plan trajectories as well as strong and soft problem goals on a plan. Nevertheless, it appears to be insufficient while trying to capture the semantics that we use for the norm lifecycle mainly due to two reasons: 1)it lacks the operator “until”, which would permit us to express the norm lifecycle (ex. a norm is violated when activated at some point and maintenance condition does not hold at some state after this and deactivating condition does not hold at any state in between) and 2)a norm can be activated and deactivated (and possibly violated) several times during the execution of a plan, something not possible to be expressed in PDDL 3.0. II. STRIPS PLANNING Our normative planning framework will extend STRIPS with additional normative elements (see Section IV) in order to allow for normative reasoning within the planning process. In this section, we briefly describe the semantics of the STRIPS formalisation [2]. Definition 1: We define F as set of fluents F = {f 1 ,f 2 ,...,f m } where fluent f i is an atomic proposition (propositional property). The state of the world is defined in terms of fluents that hold at the particular situation. We define a state to be a (possibly empty) subset of F, i.e. a state is represented by the set of fluents that are true in it. The fluents not in the state are assumed false. Each combination of fluents forms a different state, and the union of all the states is a set of states as in Definition 2. Definition 2: We define S to be the state domain (set of all states) occurring from F as S =2 f . 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology 978-0-7695-4513-4/11 $26.00 © 2011 IEEE DOI 10.1109/WI-IAT.2011.249 33 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology 978-0-7695-4513-4/11 $26.00 © 2011 IEEE DOI 10.1109/WI-IAT.2011.249 33