Agent-Based Software Engineering

This talk summarizes information from:

• Nicholas R. Jennings. On agent-based software engineering. [1] Artificial Intelligence,(117):277--296, 2000.

• Building high quality software for real-world applications is getting more difficult as the problems get more complex (no more Pong).
• Current software engineering methods
  • Objects
  • Components
  • Design Patterns
  • Software Architectures (systems of patterns).
• are not always enough because they fall short in two main ways.
  • The interactions between computational entities are too rigidly defined.
  • They lack mechanisms for representing the system's inherent organizational structure.

• Agent-oriented approaches can build complex systems better. (The Adequacy Hypothesis)
• Agent-oriented approaches will be used by the mainstream. (The Establishment Hypothesis).

Nick Jennings shows us how.

"An agent is an encapsulated computer system that is situated in some environment and that is capable of flexible, autonomous action in that environment in order to meet its design objectives."

• So, an agent is:
  1. A problem solving entity.
  2. Situated in an environment.
  3. Designed to achieve a specific purpose.
  4. Autonomous.
  5. Capable of exhibiting flexible problem solving behavior in pursuit of its design objectives.
• Therefore, agents
  • Communicate at the Knowledge Level [2].
  • Agents make run-time decisions about their interactions and respond to interactions that were not foreseen at design time.
  • When agents interact there is usually an organizational context.

• In order to tackle the complexity of software we need to use the tools of (Booch):
  • Decomposition
  • Abstraction
  • Organization (interrelationships between problem solving components).
• The characteristics of a complex system are:
  • Complexity frequently takes the form of a hierarchy--a system composed of interrelated subsystems.These are dynamic relationships.
  • We can distinguish between the interactions among subsystems and the interactions within subsystems. Complex systems are nearly decomposable.
  • The choice of primitive components is nearly arbitrary.
  • Hierarchical systems evolve more quickly than non-hierarchical ones because they have stable intermediate forms.

Where we show that agent-oriented decomposition is an effective way of partitioning the problem space of a complex system.

• Autonomy helps because:
  • It is a natural representation of complex systems.
  • Since decisions are made by autonomous entities selection can be made based on the local situation.
• Flexible interactions help because:
  • Run-time decisions about interactions means designer does not have to predict a priori all possible links, which would be impossible.
  • Managing relationships is much easier since they are done bottom-up.
• For example, manufacturing systems have been built where every part is represented by an agent whose objective is getting itself to the end of the line, having a specified set of operations performed upon it. These systems have been successful because
  • Localized decision-making means that when machines break we can route around them.
  • Since schedules are built dynamically, they can be easily altered.
  • The explicitly defined relationships between parts identify the agents that need to coordinate their actions.

Where we show that key abstractions of the agent-oriented mindset are a natural means of modeling complex systems.

• Subsystems and agent organizations are nearly the same thing.
• The interplay between subsystems and their components can be naturally viewed in terms of high level social interactions.
• Complex systems treat collections as single units. Agents systems also do the same by using roles, norms, social laws, interaction protocols for group formation, joint intentions, and teams.

Where we show that the agent-oriented philosophy for dealing with organizational relationships is appropriate for complex systems.

• Organizational constructs are first-class entities in agent systems.
• Agent systems can flexibly form, maintain, and disband organizations.
• This allows an agent system to vary the notion of a primitive component according to the needs of the observer (from agent to collections of teams).
• It also provides a variety of stable intermediate forms which are essential for the rapid development of complex systems.

• The agent-based approach can be viewed as the next step in the evolution of software engineering.
  • Languages have evolved from being based on the machine's architecture to being based on the problem domain. Agents are more common than objects.
  • The basic building blocks have evolved to have more localization and encapsulation. Agents localize purpose and action selection.
  • More re-sue. Agents enable whole subsystems and flexible interactions to be re-used. They enable the re-use of legacy code (stable intermediate forms again).
• Agents are ideal for developing software in open, networked systems. These systems need
  • Individual problem solving entities that can achieve the specified objectives.
  • The entities must be able to handle the uncertainty.
  • They need to interact with entities that were not foreseen at design time.
  • Any organizational relationships between stake-holders must be reflected in the behaviors of the entities (i.e., do what's good for me).

• The patterns and the outcomes of the interactions are inherently unpredictable. (debugging is lots of fun!)
• Predicting the behavior of the overall system based on its constituent components is extremely difficult because of emergent behaviors (but, I like emergence!)
• We can avoid these with good incentive-compatible design (game theory), using rigid organizational structures, or limiting the interactions.

• Perhaps, we should view agents at a new social level.
• Allen Newell proposed the Knowledge Level [3]. The levels are:
  1. Device
  2. Circuit
  3. Logic
  4. Combinatorial and Circuit
  5. Register-Transfer
  6. Program
  7. Knowledge
• Nick is proposing a new Social level:

  Dimension	Description	Knowledge Level	Social Level
  System	Entity to be described	(asocial) Agent	Agent organization
  Components	The system's primitive elements	Goals, Actions	Agents, Interaction channels, Dependencies, Organizational relationships
  Compositional law	How the components are assembled	Various	Roles, Organization's rules
  Behavior law	How the system's behavior depends upon its composition and components	Principle of rationality	Principle of organizational rationality
  Medium	The elements to be processed to obtain the desired behavior	Knowledge	Organization and social obligations, Means of influencing others, Means of changing organizations structures