Domain

Remarks

What is an agent?

• Legally, an agent is empowered to represent the client in activities & actions that are explicitly approved by the client
• Thus, we often encounter people as insurance agents, property agents, auction agents, company agents, etc.
• Simply put, an agent in artificial intelligence (AI) is a software that performs tasks assigned by the users
• Hence, there are information agents (gather required info.), filter agents (screen & organize data or emails), control agents (maintain operations & processes), etc.

What is an intelligent agent?

• An intelligent agent (IA) is a software agent that has the programmed capability of AI
• IA can learn (machine learning) & reason (inference) as well as operates non-linearly (adaptive, dynamic)

What is a multi-agent system?

• MAS is a system has more than one IA working either independently or co-operatively or competitively
• Peter Stone: four types of MAS can be inferred depending on the two parameters of Homogeneity & Communication, arranged below in ascending order of complexity:

1. Homogeneous + non-communicating: all IA's are duplicates of one another & all operate without interaction (identical, independent agents)
2. Homogeneous + communicating: all IA's are duplicates of one another & interact mutually (identical, interacting agents)
3. Heterogeneous + non-communicating: IA's are designed & implemented differently & operate without interaction (variable, independent agents)
4. Heterogeneous + communicating: IA's are designed & implemented differently & interact mutually (variable, interacting agents)

• Teamwork within MAS under real-time, noisy, collaborative, and adversarial environments:

1. Team-member agent architecture: flexible structure
2. Layered learning: hierarchical learning effects
3. Multi-agent reinforcement learning algorithm, namely team-partitioned, opaque-transition reinforcement learning (TPOT-RL)
4. Fully functional MAS used in Robocup soccer

What can MAS be used for?

• Intellectually, it is both exciting & challenging to create & observe a society of IA's, valuable insights and implications can be modelled & derived
• In research, MAS justify the further investment into IT and being adaptive in nature, MAS can be flexibly developed for various applications & networks (things like control, robotic soccer, artificial life, evolution, …)
• In applications, MAS has already been used for Internet and business intelligence & e-commerce; MAS has also been applied to problems involved in acting in the physical world, such as distributed traffic control, flexible manufacturing, the design of robotic systems and self-assembly of structures
• In the future, MAS can be improved upon & incorporated ubiquitously into education, entertainment, work, etc. The potential for good is large, but so is the potential for bad, the right use is crucial

Appropriate situations for MAS?

• According to Wooldridge 2002:
• Environment is open, highly dynamic, uncertain or complex: like weather conditions, economic conditions, social conditions, traffic conditions, Internet conditions, etc.
• Agents are naturally recognised or assumed: in macro or micro-economic conditions
• Distribution of data, control or expertise: such that no centralised body exists or is able to govern the distribution
• Legacy systems: technically obsolete, but functionally working systems that require agents as interfaces with external environment

Concerns & Pitfalls of MAS

• Note the following points to consider when sourcing for solutions to complex problems or designing & developing actual MAS (Wooldridge 2002 "Introduction to MAS"):
• Incomplete knowledge of agents & MAS: why do U need agents? Beware of both limitations & capabilities - seek the right balance
• Appropriateness of using MAS: is MAS the only solution? Are there alternatives? - look for effectiveness, what is effective need not be complicated like MAS
• Generic vs. customised solutions: when & how is a generic solution like MAS useful? Can a customised solution be more effective?
• Are agents a silver bullet?: Can agents solve everything? Can U balance the perceptions & arguments against the actual results? - be pragmatic
• Inherent limitations: MAS is a multi-disciplinary approach to problem-solving, hence it inherits the verified limitations of its components like software engineering, AI, multi-threading (complex class), networking, communications, etc.
• Ability to exploit the full potential of MAS when appropriate: Exploiting the software aspects (methodologies & experimentation), the concurrency (do U need to have simultaneous but independent processing), autonomy (independence with self-interest), adaptation (insensitive to changes), number of agents (too many or too few), right complexity (just sufficient for solution, not overburdened with peripheral & experimental techniques like interfaces, provers & reasoning), etc.
• MAS architecture & infrastructure: When to use off-the-shelf designs or customised ones? Do U need to build from scratch? Are there software libraries & tools available for customisation? Should U focus more on MAS architecture than the agents? How to structure the agents & MAS for simplicity, effectiveness & efficiency?

Aspects of MAS

• Two levels:
• Agent level:

1. Functional: performs specific local tasks with local objectives, regardless of external situations
2. Reasoning: performs individual decision-making
3. Strategy: sequence of actions intended to reach local objectives
4. Interfacing: to accept inputs & produce outputs to external environment

• MAS level:

1. Multiagent interaction: game theory, (dominant) strategies, Nash equilibria, dependence relations
2. Agreement: negotiations, auctions
3. Communication: languages like KQML, ontologies, coordination languages
4. Working together: cooperation (non zero-sum interactions: mutual gains possible), coordination, competition (zero-sum interactions: one gains, others loses)

Resources for MAS?

• Agent Construction Tools
• Java Programming Resources -- Java, Java, and more Java

Any plan for MAS implementation?

• User requirements: uses, needs, wants
• Specifications: objectives, scope, functions, details
• Literature review: existing research, achieved, lacking holes, critiques, niche to fill
• Skills: programming (Java, Labview, C++, Matlab), experimentation (equipment available, setup, people), AI & IA, tools (software agent builder)
• Management: progress setup & tracking, reports (templates), balance of datelines + innovations
• Synergy: obstacles to be cleared, ideas to be inspired, people to collaborate & cohere
• Operations: systems design, writing, implementation, testing, debugging & finalising
• Writing: reporting results, findings, implications & conclusions
• Finale: passing on & spawning the fruits of innovation & labour