Skip to main content

An evolutionary approach to automated norm synthesis for multi agent systems

Periodic Reporting for period 1 - EVNSMAS (An evolutionary approach to automated norm synthesis for multi agent systems)

Reporting period: 2016-06-06 to 2018-06-05

The area of multi-agent systems (MAS) is concerned with the design and implementation of systems where autonomous agents interact, usually to achieve their own goals, and sometimes to solve complex problems that cannot be solved individually. MAS are present in human societies, e.g. . (autonomous) cars in a road, mobile robots in warehouses, autonomous drones (UAVs), etc. Within the MAS domain, a key open problem is that of coordination, i.e. how to manage the interactions between agents. One of the most successful coordination approaches is that of normative sytems, that is, the use of norms that coordinate the agents by specifying what they can and cannot do. Some desirable properties for normative systems are stability and effectiveness, i.e. norms that the agents will comply with (stable), and whose compliance will successfully achieve coordination (effectiveness).

Synthesising stable and effective normative systems is a crucial problem for system designers and policy makers. For instance, with the advent of autonomous cars, it will be crucial to design norms that cars will comply with because non-compliance will be prejudicial for them; and whose compliance will avoid undesirable outcomes such as collisions. However, designing norms for MAS can be a highly complex, time consuming and error prone task, specially when the agents may have different goals and preferences. For example, autonomous cars might be made by different companies that establish different driving policies for their cars. For this reason, several approaches have been proposed for the automatic design (synthesis) of normative systems, even though it still remains an open problem.

The main goal of this project is to develop a framework for the automatic synthesis of stable and effective normative systems for system designers and policy makers. Our framework roots in the framework of Evolutionary Game Theory (EGT), which provides a mathematical framework and an algorithm for the prediction of stable strategies in MAS. Our goal is to build on the framework of EGT in order to develop:
1. A mathematical framework to model normative systems in MAS from an EGT perspective.
2. Equations and algorithms for the simulation of the evolution of norms in MAS.
3. A computational framework for the automatic synthesis of stable and effective norms.

At the end of the action, we successfully achieved the main goal of the project. We developed a framework called SENSE (Synthesis of Evolutionarily stable Normative SystEms) that employs EGT to automatically synthesise stable anf effective. SENSE takes as input descriptions of an agent population and a collection of games modelling different coordination situations, and outputs sets of stable norms that effectively coordinate these agents as required by a policy maker.
"During the first year of the project, we developed the aforementioned mathematical framework and the computational framework that allow to perform norm synthesis. We first built on classical EGT to develop a formal framework to model normative systems in a game-theoretic setting. We designed appropriate fitness functions and replicator equations to simulate the evolution of norms and their compliance. Then, we developed an architecture along with an algorithm for the synthesis of stable and effective norms. We evaluated our framework in a simulated traffic domain in which agents are autonomous cars, and the goal is to avoid collisions. We published our results in two papers in an international conference and a relevant journal of the area.

During the second year of the project, we made our framework more accessible to the general public. In our first version (the one we developed during the first year), a dedicated agent-based simulator was required in order to synthesise norms for specific domains. For instance, to synthesise traffic norms, we had to develop a traffic simulator to simulate cars' interactions. This made our framework unaccessible to those who do not have the necessary time and knowledge to develop agent-based simulators. Hence, we developed a domain-independent version of our framework in which domain knowledge can be represented as simple games in a game-theoretic setting. Our framework allows to represent complex coordination situations as combinations of simple (pairwise) agent interactions that can be co-related. We validated our framework in the traffic scenario and an example scenario of Unmanned Aerial Vehicles (UAVs). Currently, we are finishing to write a journal paper describing our newest version of the framework along with its validation in this latter scenario.

Additionally, we started a complementary line of research on the synthesis of moral norms. Synthesising moral norms is an important problem due to the advent of autonomous (self-driving) cars, UAVs (autonomous drones), Swarm Robots, etc. In human societies, desired behaviour usually aligns to some moral values that represent ""what is acceptable and what is not"". Synthesising norms aligned with the moral values of a society may have potential impact on the context of autonomous vehicles, for instance to learn which norms will be appropriate for the cars while driving in a particular city or country. During this action, we developed a framework for the synthesis of norms aligned with some moral values specified by a policy maker. We published this work in three papers in international conferences of the area. Currently, we are preparing a journal paper introducing our latest advances and a complete validation in a simulated scenario.

Our mid-term goal is to merge our line of research on moral norms with the work developed in this project, thusu providing a framework for engineering stable, effective, and moral norms."
With the realisation of this project, we made significant advances beyond the state of the art. Previously to this project, most works on norm synthesis focused on developing mechanisms whereby the agents can synthesise their own norms at runtime (norm emergence). These works required agent-based simulation in order to validate their effectiveness. Alternatively, some works focused on the off-line design of norms and developed mechanisms to synthesise effective (and sometimes stable) norms in one single situation usually modeled as a game in a game-theoretic setting. There, the typical approach is considering a MAS in which the agents can repeatedly play a game, and developing algorithms to synthesise norms that coordinate the agents in this game.

We contributed to the state of the art by developing a framework that considers multiple games that can be co-related. This enables policy makers to represent highly complex coordination situations (games) by means of combinations of more simple (pairwise) games, which significantly simplifies the necessary configuration process to run automatic norm synthesis. Moreover, our framework considers heterogenous populations with different goals and preferences, unlike most previous state-of-the-art approaches, which consider homogeneous populations with the same goals. This makes our framework more realistic than previous ones (since human societies are heterogeneous). For these reasons, we believe that our framework stands for a valuable contribution that may ease the task of designing norms for autonomous robots, cars, drones, etc.