Multiagent Systems

From Michael Wooldridge:

The complete proceedings of the AAMAS-08 Conference, held in Estoril, Portugal in May 2008, are now available FOR FREE from here.

The complete proceedings of AAMAS-07, held last year in Hawaii, are also available for free from here.

There is no catch to this, and it isn't a scam - you can access the complete proceedings for free without registering or similar. The proceedings are provided for the scientific community by the International Foundation for Autonomous Agents and Multi-Agent Systems (IFAAMAS), a not-for-profit organisation whose activities involve organising the annual AAMAS conference. The aim of IFAAMAS is to make all future AAMAS conference proceedings similarly available for free. The website of IFAAMAS is www.ifaamas.org

PRIMA (Pacific Rim International Workshop on Multi-Agents) renewed website

PRIMA is Pacific Rim workshop related to autonomous agents and multiagent systems. Though we already have several workshops in Pacific Rim countries, such as MACC (Multiagent Systems and Cooperative Computation) in Japan from 1991, and Australian Workshop on Distributed Artificial Intelligence from 1995, there is less interaction so far among the countries compared to Europe and Americas.

The aim of this workshop is to encourage activities in this field, and to bring together Pacific Rim researchers with agents and multiagent issues. Unlike usual conferences, this workshop will mainly discuss and explore scientific and practical problems as raised by the participants. Participation is thus by invitation only and is limited to professionals who have made significant contributions to the topics of the workshop. The contributions may include technical presentations, progress reports and so on.

A good example of a company that is putting multiagent techniques to work in a robotic domain is Kiva Systems. They design and sell automated robots that handle the task of finding items in a warehouse and bringing them to the front to be shipped. Their robots cooperate with each other and stay out of each other's way. Their company history summarizes their motivations well:

Founded in 2003, Kiva is focused on solving real problems in the supply chain. Founder and CEO Mick Mountz experienced the challenges of existing material handling systems firsthand while working at online grocer Webvan. The complexity of existing equipment and processes and the resulting high cost of filling orders ultimately was the downfall of Webvan. It occurred to Mick that there must be a better way to accomplish pick, pack and ship.

Mick then asked himself the simple question, "What if all the products in the warehouse could walk and talk on their own, couldn't they just come to me when I need to fill an order?" To pursue this idea, he sought the help of two experts in the area of complex multi-agent systems, Professors Peter Wurman and Raffaello D'Andrea, and together they began to develop the Kiva concept.

Today, Kiva applies the concepts of "distributed intelligence" to inventory management. Inspired by ant colonies capable of performing large and complex tasks with limited central control, the Kiva system allows inventory to organize itself, adapting to conditions as they change. The resulting solution combines store, move and sort functions into one simple system that can now deliver any item to any operator at any time.

If you are a new graduate, note that the are hiring.

The Teamcore group (teamcore.usc.edu) is focused on research on multiagent systems where multiple agents (including software agents, robots and people) may interact. We focus on fundamental research issues in Belief-Desire-Intentions (BDI) systems, in Distributed Constraint Reasoning (DCR), and in Decision Theoretic (distributed MDPs) and Game Theoretic approaches for multiagent systems. In addition to fundamental research, the group is also focused on practical implementations of their research. The most recent practical implementation is the ARMOR system deployed at the Los Angeles International Airport since August 2007. http://www.newsweek.com/id/43401

We have a new opening for a post-doctoral research associate position starting in June 2008. Research will focus on the areas of fundamental research outlined above, with emphasis on new algorithms and but also on their practical implementations/applications. Interested applicants should send their CV and have three letters of recommendation forwarded to Prof. Milind Tambe (tambe@usc.edu).

In the latest issue of IEEE Intelligent Systems, Terry Payne writes about Web Services from an Agent Perspective and tells us why agents are different from web services. Namely:

1. Agents are problem solvers.
2. Agents are pro-active.
3. Agents are goal-oriented.
4. Agents are context-aware.
5. Agents are autonomous.

Perhaps. There are some who enjoy engaging in pseudo-philosophical discussion on whether or not a piece of code is pro-active, or goal-oriented, or context-aware, etc. In the end, however, that discussion completely overlooks the real contribution of multiagent research. Multiagent research provides the higher-level algorithms for organizing complex systems. Web services provide us with the plumbing needed to get one machine to talk to another and then multiagent algorithms tell us what, exactly, those services should be. The two operate at separate abstraction levels. There should not be any confusion between the two. It's like TCP/IP versus bittorrent.

More importantly (for us), is the fact that as web services (service oriented architectures (SOA)) increase in popularity there will be a corresponding increase in the demand for the interaction algorithms and protocols developed by multiagent researchers. The budding web services ecosystem will soon demand the creation of ever more sophisticated coordination mechanisms.

Let me reiterate my point with an example. One of the most commonly used multiagent examples is the trip-planning agent. This is an agent (call it a web application if you wish) which asks you where you want to travel, when, why, and keeps a profile of your preferences. It then goes off and makes all the required purchases for you: plane, car rental, hotel, tickets to a show, conference registration, etc. We can easily envision airlines, hotels, and all the other service providers implementing a web-service backend to their website. However, this alone does not solve the problem. The agent must still figure out which specific things to buy for the user, how much to pay for them, and in what order to make the purchases (no refunds). Further, we know that the sellers themselves will also be changing prices to reflect demand. How do they make these decisions? Can two agents negotiate a deal that is better for both of them? Can we find more efficient (everyone does better) solutions by allowing more complex (multi-attribute, conditional, time sensitive) negotiations among agents? What added semantic markup do we need?

These are all very interesting questions which have not yet been fully answered. They are also the questions that multiagent systems research tries to answer. Not, “How do I make an online purchase from the Marriot hotel?”, but “What protocol and behaviors must all these agents implement in order to make trip-planning possible and incentive-compatible for all parties involved?”

New Scientist has an interesting article on shape shifting robot forms from magnetic swarm which describes multi-robot (swarm robotics) research going on at CMU and other places. The video is pretty neat too:

I also noticed this quote:

But software, not hardware, may be the biggest challenge facing researchers working on swarms of robots, he [Alan Winfield] says: "Right now we just don't know how to design a system that produces complex overall behaviours from a group of simple agents."

Luckily for them, that is exactly the problem that multiagent researchers have been tackling for over a decade. Unluckily, it is a hard problem to solve, at least for some definitions of "complex ".

Xaitment has just announced AI libraries for games. From their press release:

xaitment GmbH, one of the leading developers and service providers of artificial intelligence for the games and simulation industries, announced today that it has launched five new artificial intelligence (AI) modules ranging from standard to high-level AI functionality. The modules, which position the xaitEngine as one of the most flexible AI solutions on the market, will be shown at GDC 2008 in the German Pavilion (North Hall, Booth 6105).

The xaitEngine drives the predictable, and unpredictable, behaviors of computer-generated agents, and computer-operated applications and machines. It has primarily been used in interactive entertainment, though its applications are limitless - from the simulation of intelligent drivers in racing games, to resource management in manufacturing.

The xaitEngine handles all standard AI functionality, including pathfinding and simple AI behavior patterns such as movement. But while most other AI vendors stop there, the xaitEngine also handles more advanced AI. Such high-level AI can be found in the realistic interaction of the AI with its environment or the emotional intelligence of non-player characters (NPCs) in a game.

"Artificial intelligence has been offered in games for awhile," notes Dr. Andreas Gerber, CEO of xaitment. "But truly lifelike AI that offers emotional behaviors, autonomous actions and humanistic unpredictability, is something that almost no company has been able to provide, until now. What's more, we've been able to create our solution in a flexible manner that gives developers options so they are not locked into buying more than they need to enrich their own AI for their game."

As the marginal financial returns on graphical improvements decrease and the focus shifts over to playability, notice the popularity of the wii against the ps3 and xbox 360, we can expect to see more and more interest in using AI and multiagent techniques in games.

To me, the most interesting aspect of this development is that users will not want to battle infallibly intelligent opponents. That's boring. Instead the agents will need to incorporate emotional models, models of human-like non-rational behaviors (cf. behavioral Economics and Sociology for models), etc. They will also need to do this within a multi-player, thus multiagent, environment. In short, multiagent negotiation and decision-making techniques will find widespread adoption in the gaming domain.

The people at xaitment are betting on this, note their mission:

A spin-off of the world renowned German Research Center for Artificial Intelligence (DFKI), xaitment was founded in 2004 with the mission to create lifelike AI for games and simulations. Their mission led to the development of the xaitEngine, a highly customizable and highly modular multi-agent system that enables bots to learn from their mistakes, coordinate activities, compete with each other and achieve their goals with uncanny realism.

The NSF has announced a new center for autonomic computing. You can check out some of their presentations. The NSF blurb tells us that

An autonomic computing system is any system that is designed to function with minimal management even as conditions and users change, according to Dr. José Fortes, director of the new center at the CAC's University of Florida site. Autonomic computing algorithms, Fortes says, can greatly reduce the growing costs of administrating computer systems and protect against loss of service in systems performing critical functions, including those managing power grids, stock markets, and hospital networks. They can also greatly improve the speed and efficiency of complex systems that utilize a large number of hardware and software components. Autonomic behaviors are collectively known as "self-*" behaviors. "For instance, a system that stores secure information could use a self-protecting algorithm to detect and mount a defense against attacks," says Dr. Salim Hariri, director of the CAC's University of Arizona site. Similarly, a system that provides critical services could use a self-healing algorithm to identify and recover from disruptions triggered by hardware and/or software failures, Hariri says.

Of course, the idea of writing software that adapts to is not new: a load-balancing router for web servers adapts to varying user loads and computer crashes, a RAID5 system adapts to read-write errors and hard drive crashes. But, they seems to propose to build even more self monitoring software.

Multiagent systems provide a unique perspective to this system design problem. We can view each computer or system as a selfish agent trying to maximize its local utility then develop negotiation protocols for the agents to come to agreements about resource utilization. The system manager would then be left with the job of assigning utility values. For example, he could say that fast response time to the users is more important that volume of users (if, say, he wants to provide all users of his website with either a good experience or a 500 error page, instead of giving everyone a bad experience). The servers would then negotiate their bandwidth and database usage accordingly.

The UMBC software agents mailing list is one of the oldest resources of information on agents and multiagent systems. It was started in 1994 by Ray Johnson, then at the Lockheed Palo Alto AI Center and moved to UMBC in 1996. This week it upgraded its support infrastructure from Majordomo to GNU Mailman. Majordomo represented the state of the art for mailing list software in 1996, but development stopped sometime around 2001. Moving to Mailman will make it easier for us to maintain the list and let the ~2000 subscribers manage a wider range of their subscription options.

Topics of interest for the agents mailing list include: agent architectures, agent communication, agent learning, evolution, and adaptation, agent norms and trust, agent ontologies, agent-oriented software engineering, autonomic computing, autonomy, believable agents, cognitive agent models, cooperative distributed problem solving, electronic markets and institutions, embodied agents, emergent behavior, ethical and legal issues, FIPA standards, formal agent models, interface agents, MAS planning and learning, mechanism design, auctions, and game theory, mobile agents, multiagent systems, pervasive computing, agent-oriented robotics, simulation systems, and standardization efforts.

The agents list welcomes new subscribers who can sign up at the software agents mailing list page.

Post-doctoral Research Associate Position The Teamcore Research Group Computer Science Department University of Southern California Dec 2007

The Teamcore group (teamcore.usc.edu) is focused on research on multiagent systems where multiple agents (including software agents, robots and people) may interact. We focus on fundamental research issues in Belief-Desire-Intentions (BDI) systems, in Distributed Constraint Reasoning (DCR), and in Decision Theoretic distributed MDPs) and Game Theoretic approaches for multiagent systems. In addition to fundamental research, the group is also focused on practical implementations of their research. The most recent practical implementation is the ARMOR system deployed at the Los Angeles International Airport. http://www.newsweek.com/id/43401

We have an opening for a post-doctoral research associate position starting in March 2008 (and possibly earlier). Research will focus on the areas of fundamental research outlined above, with emphasis on new algorithms and but also on their practical implementations/applications. Interested applicants should send their CV and have three letters of recommendation forwarded to Prof. Milind Tambe (tambe at: usc.edu).