In this section, we discuss some of the most fundamental characteristics of an agent-oriented approach to solution development, which might be different from existing paradigms and methodologies that you may be used to. We hope that with this, we can guide you towards the right mindset when designing your own agent-based solutions to real world problems.
Multi-Agent Systems (MAS) are inherently decentralised. The vision is of an environment in which every agent is able to directly connect with everyone else and interact with them without having to rely on a third party acting as an intermediary or match-maker. This is in direct contrast to centralised systems in which a single entity is the central point of authority, through which all interactions happen. The conventional client-server model is an example of a centralized architecture where clients interact with one another regarding specific services (e.g. communication, commerce) only through a server.
This is not to say that facilitators and middlemen have no place in a multi-agent system; rather it is the 'commanding reliance on middlemen' that MAS rejects.
Division of responsibilities: In a decentralised system, every agent is equally privileged, and (in principle) should be able to interact with any other agent. The idea is very much aligned with the peer-to-peer paradigm, in which it is the voluntary participation and contribution of the peers that create the infrastructure. Therefore, in a decentralised system, there is no central 'enforcer'. This means all the work that would typically fall under the responsibilities of a central entity must be performed by individual parties in a decentralised system. Blockchain-based cryptocurrencies are a good example of this. A notable characteristic of cryptocurrencies is the absence of central trusted entities (e.g. banks). But this in turn means that most security precautions related to the handling of digital assets and the execution of transactions are the responsibility of individuals.
Decentralisation vs distribution: It is important to emphasise that by decentralisation we do not mean distribution; although multi-agent systems typically do also tend to be distributed. A distributed system is one whose components are physically located in different places and connected over a network. A fully centralised system, owned and operated by a single entity, may in fact be highly distributed. Google or Microsoft's cloud infrastructure are examples of this, where their components are distributed across the globe yet designed to work together harmoniously and function in unison. Decentralisation on the other hand refers to a system whose components may be owned, operated, and managed by different stakeholders, each with their own personal objectives, interests, and preferences which may not necessarily be aligned with one another or the system itself. Therefore, distribution refers to the physical placement of a system's components, whereas decentralisation refers to a) the diversity of ownership and control over a system's constituents, and b) the absence of central authorities between them.
Example: To better illustrate the distinction between centralised and decentralised systems, consider another example: search and discoverability in a commerce environment. In a centralised system (say Amazon), there is a single search service -- provided, owned and run by the commerce company itself -- which takes care of all search-related functionality for every product within their domain. So to be discoverable in this system, all sellers must register their products with this particular service. However in a decentralised system, there may not necessarily be a single search service provider. There may be multiple such services, run by different, perhaps competing entities. Each seller has the freedom to register with (i.e. make themselves known to) one or a handful of services. On the buyers side, the more services they contact and query, the higher their chances of finding the product they are looking for.
As discussed above, the notion of decentralisation extends as far as ownership and control. Therefore, the different components that make up a decentralised system may each be owned by a different entity, designed according to very different principles and standards, with heterogeneous software and hardware, and each with internal objectives that may be fundamentally inconsistent, worse yet contradictory, with those of others.
As such, a distinctive characteristic of a multi-agent environment, is that it is inhabited by more than one agent (as the name suggests), where each agent may be owned potentially by a different stakeholder (individual, company, government). Since by design, each agent represents and looks after the interests of its owner(s), and because different stakeholders may have unaligned, conflicting, or contradictory interests, it is very common to have multi-agent systems in which the agents' objectives, values and preferences are unaligned, conflicting, or contradictory.
In practice: There are practical implications that follow from the above when it comes to designing an agent. For example, it is not rational for an agent to automatically rely on the information it receives from other agents. The information could be:
- Incomplete: what is unrevealed may have been deemed private for strategic reasons.
- Uncertain: it may be the result of an inaccurate prediction.
- Incorrect: it could be an outright lie, due to the adversarial nature of the environment.
Therefore one can argue, that there is a degree of uncertainty attached to almost all information an agent receives or infers in a multi-agent system. It wouldn't then be illogical for an agent to take a sceptical approach: treating everything as uncertain, unless proved otherwise.
The conflicting nature of multi-agent systems, consisting of self-interested autonomous agents, points to asynchronisation as the preferred method of designing and managing processes and interactions.
Synchronisation vs asynchronisation: In general, asynchronisation refers to the decoupling of events that do interact with one another but do not occur at predetermined intervals, not necessarily relying on each other's existence to function. This is in contrast with synchronous systems in which processes are aware of one another, where one's execution depends in some way on the other.
Asynchronisation in MAS: In the context of multi-agent systems, the decentralised and potentially conflicting nature of the environment creates uncertainty over the behaviour of the whole system, in particular of other agents. For example, suppose an agent
i sends a message requesting some resources from an agent
j. Since MAS often tends to be distributed, there is the usual uncertainties with communication over a network:
j may never receive
i's request, or may receive it after a long delay. Furthermore,
j could receive the request in time and respond immediately, but as mentioned in the last section, its answer might be incomplete (gives only some of the requested resources), uncertain (promises to give the resources, but cannot be fully trusted), or incorrect (sends a wrong resource). In addition, since agents are self-interested,
j may decide to reply much later, to the point that the resource is no longer useful to agent
j may simply decide not to respond at all. There might be a myriad of reasons why it may choose to do that; it could be because
j assigns a low priority to answering
i over its other tasks. But that's beside the point. The take away is that agents' autonomy strongly influences what can be expected of them, and of an environment inhabited by them. As such, developing for a system whose constituents are autonomous, e.g. agents in a multi-agent system, is fundamentally different from one whose constituents aren't, e.g. objects in an object-oriented system.
Objects vs agents: In object-oriented systems, objects are entities that encapsulate state and perform actions, i.e. call methods, on this state. In object-oriented languages, like C++ and Java, it is common practice to declare methods as public, so they can be invoked by other objects in the system whenever they wish. This implies that an object does not control its own behaviour. If an object’s method is public, the object has no control over whether or not that method is executed.
We cannot take for granted that an agent
j will execute an action (the equivalent of a method in object-oriented systems) just because another agent
i wants it to; this action may not be in the best interests of agent
j. So we do not think of agents as invoking methods on one another, rather as requesting actions. If
j to perform an action, then
j may or may not perform the action. It may choose to do it later or do it in exchange for something. The locus of control is therefore different in object-oriented and agent-oriented systems. In the former, the decision lies with the object invoking the method, whereas in the latter, the decision lies with the agent receiving the request. This distinction could be summarised by the following slogan (from An Introduction to MultiAgent Systems by Michael Wooldridge):
objects do it for free; agents do it because they want to.
All of this makes asynchronisation the preferred method for designing agent processes and interactions. An agent's interactions should be independent of each other, as much as possible, and of the agent's decision making processes and actions. This means the success or failure of, or delay in any single interaction does not block the agent's other tasks.
Closely related with the discussion of asynchronicity, is the idea that in multi-agent systems, time is not a universally agreed notion. Agents may not necessarily share the same clock and this fact must be taken into account when designing agent-based systems. For example, you cannot necessarily expect agents to synchronise their behaviour according to time (e.g. perform a certain task at a time
Another related issue, is that unlike some agent-based simulation (ABS) systems where there is a global tick rate for all agents, in AEA-based systems tick rates may be different for different agents. This is due to the fundamental difference that ABS systems control some aspects of all of their agents' executions while in AEA-based systems, agents are truly decoupled from one another - most likely distributed and running on different machines and networks - and there is absolutely no central unit that moderates any aspect of their behaviour.
Complex, Incomplete, Inconsistent and Uncertain¶
The fourth characteristic(s) relate to the environment in which agents are expected to operate in, and these have been mentioned a number of times in the previous sections.
The environment agents are suited for typically tend to be complex, to the point that it is usually impossible for any single agent to perceive the whole of the environment on its own. This means that at any point in time, any agent has a limited knowledge about the state of the environment. In other words, the agents;' information tend to be incomplete due to the complexity and sophistication of the world in which they reside.
Consider an agent which represents a driverless vehicle. The complexity of the problem of driving on the road makes it impossible for a single vehicle to have an accurate and up-to-date knowledge of the overall state of the world . This means that an agent's model of the world is at best uncertain. For instance, the vehicle, through its sensor may detect green light at a junction, and by being aware of what it means, it may infer that it is safe to cross a junction. However, that simply may not be true as another car in the opposite direction may still cross the junction violating their red light. Therefore, there is uncertainty associated with the knowledge "it is safe to cross the road because the light is green", and the agent must recognise that.
Furthermore, the often conflicting nature of the environment means information obtained from multiple sources (agents) may be inconsistent. Again, this must be taken into consideration when designing an agent which is expected to operate successfully in a potentially conflicting environment.
- Wooldridge, M. (2009). An Introduction to MultiAgent Systems. Wiley, Second edition.
- Shoham, Y. and Leyton-Brown, K. (2008). Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations. Cambridge University Press