Normative Cooperation among Autonomous Agents
Abstract
The ability of autonomous agents to cooperate is becoming a critical issue in a world in which transport, distribution, manufacture,and warfare are increasingly carried out by intelligent machines.In industrial automation, cooperation is typically achieved using a, central controller or supervisor, that dictates the order of actions and events to ensure effective synchronisation of the actions, of individual robots. This works well in a predictable and repetitive environment. However, centralised cooperation requires that e,ach of the robots/agents has been designed to accept the commands of the central controller. In a heterogeneous environment with an,unpredictable variety of agents (for example autonomous vehicles in a city, or robot soldiers in a war zone) where there is a variab,le and possibly hostile environment, this approach is no longer practical. The central supervisor is a single point of failure and,the necessity of continuously communicating with each agent is extremely vulnerable to network delays, jamming and failure. The aim,of this project is to develop novel techniques to allow effective cooperation between heterogeneous autonomous agents that do not re,ly on central control or communication but instead utilise norms. Human society relies on many norms, from the legal system and traf,fic regulations, to social conventions, such as good manners. Norms can be seen as standards of behaviour that promote effective co,operation. For example, if everyone drives on the right (or left) then collisions are avoided. Critically normative cooperation requ,ires only that agents are able to observe each other act in the environment.The project will investigate how norms can be effectivel,y learned. Autonomous robots and other agents will first learn to achieve their individual goals using Deep Reinforcement Learning i,n a Multi-Agent Reinforcement Learning simulation. To achieve collaborative goals and to optimise cooperation a set of norms will th,en be learned by the agent community. This does not require any homogeneity in the skills learned to achieve individual goals, which, may differ between classes of agent. Neither does it require any predefined command structure or communication channels between age,nts. All that is needed is a local ability to observe other agents and the shared environment. Learning norms is thus analogous to l,earning how to cooperate effectively in a team, once basic task skills have been gained. The project will develop formalisms and alg,orithms that support representing and reasoning about sets of norms. Specifically, we will investigate: 1) approaches to allow forma,l verification of learned norms, i.e., that a norm, if followed by the agents, is guaranteed to result in cooperative behaviour; and, 2) approaches to the formal specification of desired cooperative behaviours that guarantee learning will converge to the specified,behaviour.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 14, 2022
- Source ID
- N629092212005
Entities
People
- Brian Logan
Organizations
- Office of Naval Research
- United States Navy
- University of Nottingham