This position paper summarizes the contributions resulting from the short term scientific mission (STSM) of the first author, from University Rey Juan Carlos, to the Kings College London. During this scientific mission, we have studied the impact of norms on the sequential decision making of rational agents, and how they can be modelled and quantitatively assessed. For this purpose, we put forward a framework based on Markov Decision Processes that allows for explicitly representing norms and sanctions. In order to illustrate the usage of the framework, we model an example in a simulated aerospace aftermarket ruled by electronic contracts in terms of norms and sanctions, and present algorithms for reasoning about it.
There have been several proposals for enabling agents to reason about norms using the Belief-Desire-Intention (BDI) paradigm, acknowledging the advantages of norms as a means to condition agent behavior. However, the emphasis in these proposals has been on the feasibility of such normative reasoning rather than on the ability to quantitatively reason about agent performance, especially in the presence of uncertainty. Markov Decision Processes (MDPs) are well suited for quantitatively assess the agent performance under uncertainty, but so far, this framework has never been exploited to reason about norms. There have been some proposals to tackle coordination problems in MDPs, but none of these proposals have employed solutions based on normative systems.
This position paper summarizes the contributions resulting from the short term scientific mission (STSM) of the first author, from University Rey Juan Carlos, to the Kings College London. In this scientific mission, we propose a solution to model normaware rational agents that combines the native strengths of the MDPs and norms in order to cover their individual limitations. We combine, on one hand, the analytic advantages of MDPs to model the domain of interaction and agent decision-making under uncertainty, and, on the other, the use of norms to coordinate autonomous agents.
The contribution resulting from this STSM is threefold3:
(i) An extension of the MDPs, referred to as Normative Markov Decision Processes
(NMDPs), that explicitly represent norms, sanctions and detection probabilities of
norm violations. The MDP framework has been extended by an additional
component that denotes a set of norms ruling system. A norm is described as a set of
states that are prohibited or obliged for the addressee agent, and a sanction
consists of a penalty and an enforced state-transition aiming at updating the current
state of this agent. The underlying intention of these sanctions is to punish the
transgressors by decreasing their utility, and moving them from violating states to
states where the norms are obeyed and/or their capabilities are constrained.
(ii) An algorithm for identifying which states of the process violate which norms, and
a technique to calculate the risk of violating a set of norms by adopting a particular
policy. In this respect, the focus of the approach is on the norms within electronic
contracts rather than any other concern.
(iii) To illustrate the application of our approach, we developed a case study in a
simulated aerospace aftermarket [
As future work we intend to study the relation between the minimization of norm violations, utility maximization and the agents’ reputation. Before signing contracts, the signatories take into account their past experiences, which provide data to construct the reputation of the agents. Our second research direction aims at generating detailed explanations of norm violations within the NMDP framework, identifying their direct and indirect causes as well as the involved agents.
This research is supported by the AT COST Action (IC0801), Spanish Ministry of
Science and Innovation through grants TIN2009-13839-C03-02 (co-funded by Plan E) and
CSD2007-0022 (CONSOLIDER-INGENIO 2010).
3 For the sake of space, this position paper does not include technical details, such as formal
definitions and algorithms. The complete research is published in Fagundes et. al [