Multiagent Systems
Engineering (MaSE)
Robotics & Automation Engineering
Multiagent systems topology
Example of MaSE
A multiagent systems engineering approach to a large supply chain would seek to identify the specific capabilities of the supply chain components, which could include warehouses, inventory, and store clerks, and how best to optimize those capabilities to maximize overall supply chain efficiency. These characteristics can be identified through consideration of the environment they are expected to operate within and the tasks they will be expected to perform.
A classic example of a poorly matched agent with its task and environment is that of the spider in the bathtub. A spider is well adapted to exist in highly unconstrained environments such as bushes, trees, rocks, even indoor walls; however, the environment of a bathtub, with its high, smooth walls, provides an insurmountable challenge. Similarly, the physical characteristics of an engineered agent must be well matched to its environment and operational function.
Example of a Warehousing Logistic Control Design
Automated guided vehicles (AGVs) are widely used in automated warehousing functions; however, these machines operate at very low levels of autonomy – in highly constrained environments and under close monitoring. An AGV-like machine may be identified as part of a multiagent system analysis, but its behavioral modes would be such that the AGV could operate without external guidance and monitoring, in a highly unconstrained environment, able to adapt quickly to changes in its environment or goal. In this case, the AGV is an agent within the automated warehouse multiagent system. However, we may also want to coordinate the behaviors of individual AGVs as a subsystem. The MaSE approach then seeks to determine the individual AGV characteristics and behaviors that will lead to a coordination of all AGVs toward a predefined system-level goal. For instance, the AGVs, as a group, will behave differently if they are able to actively communicate with each other, as opposed to sensing the activities of other AGVs through changes in the environment. The group will also behave differently if an individual AGV seeks out other AGVs, rather than avoids other AGVs, etc.
The field of multiagent systems has recently begun to shift into a new phase – from analysis and characterization to design. Multiagent systems engineering (MaSE) is a methodology by which knowledge of the behavioral mechanisms of multiagent systems is used to design systems that are composed of intelligent, autonomous individual components (agents) and encompass decentralized control architectures. The multiagent approach provides an opportunity to approach complex engineering challenges for which a traditional solution is extremely complex or impossible.
The Multiagent Systems Engineering Approach
The MaSE approach seeks to specify the physical and behavioral characteristics of agents in such a way that the system-level meta-behaviors emerge from the spatiotemporal agent dynamics.
Aspects of the MaSE process
The process of identifying agent characteristics is pursued within a broader context of system-level design where the design of individual agents is matched with the desired behavior of the system. This approach is iterative and is often referred to as "top-down and bottom-up." The desired system behavior must first be understood. From this, the types of agents and their specific characteristics can be identified.
In this process, feedback is essential to validate that the agent design will produce the desired system effect. This feedback often takes the form of simulation, where the effects of agent characteristics on the overall system behavior can be demonstrated. Agent characteristics can be modified based on this feedback, and the simulation can be iterated to identify an optimal set of agent characteristics for the specific desired system behaviors.
Engineering Application of Multiagent Systems
One way the principles of multiagent systems might be applied to an engineered system is in manufacturing and warehousing logistics control design. These systems can be very complex in their organizational and operational structures and are traditionally controlled using top-down (hierarchical) techniques. This approach, however, introduces limits on the functional complexity and operational efficiency of the system. Hierarchical control also introduces operational fragility because the system is unable to adapt to anomalous or unforeseen events.
In terms of a warehousing process, MaSE may be used to identify equipment and process requirements in terms of physical capabilities and behavioral modes.
Summary of Multiagent Systems Engineering
This paradigm represents a shift away from traditional "systems engineering," and it is an active area of academic research. Although the concepts of agent-based modeling and multiagent systems are not new, their application to physical engineered systems is.
Complexity researchers, engineers, and scientists from many disciplines are converging toward a fundamentally new approach to the science and engineering of large, multivariable systems. Rather than maintain system intelligence externally and centrally located, the multiagent systems engineering approach distributes the intelligence throughout the system components. Rather than anticipate and explicitly describe all possible operational permutations the system may encounter, MaSE seeks to create a system that is capable of reacting in real time to any number of possible conditions. This is the essence of autonomy. A multiagent system approach could, for example, prepare a planetary explorer rover to autonomously explore the surface or atmosphere of a planet, seeking targets, avoiding obstacles, and making decisions in real-time based on the environmental information available to it. A variation of this is a planetary exploration team of simple explorer rovers, perhaps as simple as a large number of stationary sensor banks, that are scattered across the planetary surface by a delivery vehicle. Each unit may have only limited range or sensory capabilities, but taken together, they could potentially provide detailed topological or surface chemistry information. Such a group of stationary sensor banks could also provide detailed temporal information, such as temperature fluctuations.
Autonomous vehicles in manufacturing logistics
That is the challenge with designing such systems. Some questions we must ask, which cannot be answered through hierarchical approaches, are:
- What sensors do we give the individuals so the entire group performs as desired?
- Are the individuals a homogeneous or heterogeneous group?
- Do they actively communicate with each other using RF or pulsed LEDs?
- If they are mobile, do they seek out others like themselves, or avoid?
- Do they seek to acquire and maintain a specific distance between themselves, a field potential?
- In a hostile terrestrial environment, such as a combat zone or within a search-and-rescue mission, should the individuals "swarm," or gather, toward a target of interest, or should they avoid the target?
- How will the system of individuals behave differently if we program them to act sooner than later on a given decision?
- What behaviors can we give our mobile planetary explorer rovers so they don't fall into a deep ravine en masse like a flocking group of electromechanical, interplanetary lemmings?
The answers to these questions are not straightforward, and in general cannot be addressed for an unconstrained environment using hierarchical methods. The intelligence must be distributed to the individuals within the system, and the system as a whole must be considered far more than the sum of its parts. SwRI has the technical expertise to address these design issues using the tools and techniques of multiagent systems engineering.
Related Terminology
agent-based systems • multiagent systems • distributed intelligence • decentralized control • architectures • multiagent systems engineering • engineered system
