The new Scientist reports on a recent paper by CMU psychologist Don Moore that shows that people prefer advice from confident sources even when they have a poor track record.

Moore argues that in competitive situations, this can drive those offering advice to increasingly exaggerate how sure they are. And it spells bad news for scientists who try to be honest about gaps in their knowledge.
…
In Moore’s experiment, volunteers were given cash for correctly guessing the weight of people from their photographs. In each of the eight rounds of the study, the guessers bought advice from one of four other volunteers. The guessers could see in advance how confident each of these advisers was (see table), but not which weights they had opted for.
…
Describing his work at an Association for Psychological Science meeting in San Francisco last month, Moore said that following the advice of the most confident person often makes sense, as there is evidence that precision and expertise do tend to go hand in hand. For example, people give a narrower range of answers when asked about subjects with which they are more familiar”

Why aren’t we better at recognizing cover-confidence? There must be some evolutionary fitness in this, at least for humans. There can be a big penalty in indecision or vacillation. I wonder if we will see the same phenomenon in systems of cooperating autonomous agents?

Here’s the paper:

Joseph R. Radzevick and Don A. Moore, Competing To Be Certain (But Wrong): Social Pressure and Overprecision in Judgment, 21st Annual Convention of the Association for Psychological Science, May 2009.

Overprecision in judgment is both the most robust and the least understood form of overconfidence. Overly precise judgments claim more certainty than is objectively warranted. In this paper, we investigate whether the competitive social pressure of a market contributes to overprecision among those competing for influence. We find evidence that markets do indeed exacerbate overprecision. This evidence comes from two experiments in which advisors attempt to sell their advice. In the first experiment, advisors must compete with other advice sellers. In the second, advisors and decision makers are paired. Overprecision exists in both studies, and it helps advisors’ sell their advice. However, the market also exacerbates overprecision. We discuss the strategic implications of these results.

Google wave looks interesting. Google describes it as “a new tool for communication and collaboration on the web” and it’s a funny mix of email, instant messaging, wikis, and Facebook wall interactions. Or maybe IRC for the new century. This is from a post, Went Walkabout. Brought back Google Wave, on the Google blog.

“A “wave” is equal parts conversation and document, where people can communicate and work together with richly formatted text, photos, videos, maps, and more. Here’s how it works: In Google Wave you create a wave and add people to it. Everyone on your wave can use richly formatted text, photos, gadgets, and even feeds from other sources on the web. They can insert a reply or edit the wave directly. It’s concurrent rich-text editing, where you see on your screen nearly instantly what your fellow collaborators are typing in your wave. That means Google Wave is just as well suited for quick messages as for persistent content — it allows for both collaboration and communication. You can also use “playback” to rewind the wave and see how it evolved.”

Google Wave is not available yet, but you can sign up to be notified when it’s launched.

Here’s a random thought. Our models for communication in multiagent systems (e.g., KQML and FIPA) were informed by if not based on email and, to a lesser degree, IM. If Wave is a useful new communication model for humans, does it have a counterpart for software agents? If so, I suspect that ideas from the Semantic Web will be useful to provide a “rich content” for agents.

For more views, see posts by o’reilly, techcrunch, BusinessWeek and Gabor Cselle.

We’ve been working to get the dissertaions of our recent PhD graduates online. The latest one is Olga Ratsimor’s 2007 dissertation on bartering for goods and services in a mobile or pervasive environments. Here is the citation and abstract. You can click through on the title to get a pdf copy of the dissertation.

Olga Vladi Ratsimor, Opportunistic Bartering of Digital Goods and Services in Pervasive Environments, Ph.D. Dissertation in Computer Science, University of Maryland, Baltimore County, August 2007.

The vision of mobile personal devices querying peers in their environment for information such as local restaurant recommendations or directions to the closest gas station, or traffic and weather updates has long been a goal of the pervasive research community. However, considering the diversity and the personal nature of devices participating in pervasive environments it is not feasible to assume that these interactions and collaborations will take place with out economically-driven motivating incentives.

This dissertation presents a novel bartering communication model that provides an underlying framework for incentives for collaborations in mobile pervasive environments by supporting opportunistic serendipitous peer-to-peer bartering for digital goods such as ring tones, MP3’s and podcasts.

To demonstrate viability and advantages of this innovative bartering approach, we compare and contrast the performances of two conventional, frequently employed, peer-to-peer interaction approaches namely Altruists and FreeRiders against two collaborative strategies that employ the Double Coincidence of Wants paradigm from the domain of barter exchanges. In particular, we present our communication framework that represents these collaborative strategies through a set of interaction policies that reflect these strategies. Furthermore, we present a set of results from our in-depth simulation studies that compare these strategies.

We examine the operation of the nodes employing our framework and executing these four distinct strategies and specifically, we compare the performances of the nodes executing these strategies in homogeneous and heterogeneous networks of mobile devices. We also examine the effects of adding InfoStations to these networks. For each of the strategies, we observe levels of gains and losses that nodes experience as result of collaborative digital good exchanges. We also evaluate communication overhead that nodes incur while looking for possible collaborative exchange. Furthermore, this dissertation offers an in-depth study of the swarm-like inter-strategy dynamics in heterogeneous networks populated with diverse nodes displaying varying levels of collaborative interaction attitudes. Further, the bartering framework is extended by incorporating value-sensitive bartering models that incorporate digital goods and content valuations into the bartering exchange process. In addition, the bartering model is extended by integration of socially influenced collaborative interaction that exploit role based social relationships between mobile peers that populate dynamic mobile environments.

Taken as a whole, the novel research work presented in this dissertation offers the first comprehensive effort that employs and models opportunistic bartering-based collaborative methodology in the context of serendipitous encounters in dynamic mobile peer-to-peer pervasive environments where mobile entities negotiate and exchange digital goods and content.

I guess this is the ultimate question for a Turing Test. At least for this Fall.

Reporter Will Pavia of The Times was one of the judges a the 2008 Loebner Prize competition last week. In a story in The Times yesterday, Machine takes on man at mass Turing Test, he revealed his question that gave away one of the cold, lifeless, mechanical bots.

“The other correspondent was undoubtedly a robot. I asked it for its opinion on Sarah Palin, and it replied: ‘Sorry, don’t know her.’ No sentient being could possibly answer in this way.”

Of course, this could have been an ironic response from a clever person who was mocking VP candidate Palin’s stock question of “Who is Barack Obama?”.

(spotted on Languae Log)

On Sunday October 12, six computer chatterbots will sit down with six human judges at the University of Reading and try to convince them that they are not machines, but humans. The winner might take away the grand Loebner Prize worth $100,000. The Loebner Prize competition is a modified and simplified Turing test intended as a measure of machine intelligence. Here’s how Wikipedia describes it.

“The Loebner Prize is an annual competition that awards prizes to the Chatterbot considered by the judges to be the most humanlike of those entered. The format of the competition is that of a standard Turing test. In the Loebner Prize, as in a Turing test, a human judge is faced with two computer screens. One is under the control of a computer, the other is under the control of a human. The judge poses questions to the two screens and receives answers. Based upon the answers, the judge must decide which screen is controlled by the human and which is controlled by the computer program.”

This year, the competition is taking place ar Reading under the direction of Professor Kevin Warwick. The thirteen initial entries which have been reduced to six finalists.

The competition was started in 1990 by Hugh Loebner, who put up a set of cash prizes, including one worth $100,000 for the “first chatterbot that judges cannot distinguish from a real human in a Turing test that includes deciphering and understanding text, visual, and auditory input.” A fact of local interest is that Hugh Loebner worked at UMBC as the assistant director of computing in the 1980s. He left UMBC to run his family’s business, which at the time was doing well manufacturing roll-up disco dance floors for parties.

Over the years the Loebner prize competitions has come under considerable criticism from the AI research community. A common option among AI researchers is that the competition is more about publicity than science and encourages people to try to do well by exploiting tricks and competition-specific strategies rather than work on the fundamental problems underlying the development of intelligent machines. This article in Salon, Artificial stupidity, summarizes the positions.

Here are some stories on the 2008 Loebner Prize competition in the press: ‘Intelligent’ computers put to the test’, Invasion of ‘human’ robots and Artificial Conversational Entities: Can A Machine Act Human and Be Given ‘Rights’?.

Jim Odell, the acting chair of the FIPA IEEE Computer Society standards committee, recently sent out an update to the members on current activities.

“FIPA is currently working with the OMG on agent standardization, including an SOA standard that includes agents (SOA-Pro) and an Agent Metamodel and Profile (AMP). The Agent Metamodel and Profile RFP has many companies that are participating, including (but not limited to): HP, Unisys, CSC, Deere & Co, Thales, Metropolitan Life, SINTEF, and DFKI. If you are interested in participating, please let me know.

Any comments on the Agent Metamodel and Profile (AMP) RFP are welcomed. (The above companies and RMIT have already submitted their suggestions. The current release can be downloaded from: http://www.omg.org/cgi-bin/doc?ad/2008-06-02”

The OMG Agent Platform Special Interest Group page maintains links to documents about these emerging agent standards.

What determines our behavior or beliefs? Are we influenced by people who are the well-known and popular leaders — political, social, religious — in our society or by the few hundred people that are in our immediate social network — family, friends and co-workers. It’s reasonable to assume that it varies by domain or topic, with your music preferences falling in the first category and your spiritual orientation in the second.

Paul Ormerod and Greg Wiltshire have a preprint of a paper ‘Binge’ drinking in the UK: a social network phenomenon (pdf) that reports on a study that the binge drinking phenomenon seems to spread through “small world” social networks rather than by imitating influentials in a “scale free” network

“We analyse the recent rapid growth of ‘binge’ drinking in the UK. This means the consumption of large amounts of alcohol, especially by young people, leading to serious anti-social and criminal behaviour in urban centres. We show how a simple agent-based model, based on binary choice with externalities, combined with a small amount of survey data can explain the phenomenon. We show that the increase in binge drinking is a fashion-related phenomenon, with imitative behaviour spreading across social networks. The results show that a small world network, rather than a random or scale free, offers the best description of the key aspects of the data.”

It’s fascinating that with the right data, simulation models can help to answer such questions.

AP reports that an RPI group is developing a robot for Second Life, Researchers teach ‘Second Life’ avatar to think. Actually, it’s a robot with the brain of a four-year old pre-schooler.

“Edd Hifeng barely merits a second glance in “Second Life.” A steel-gray robot with lanky limbs and linebacker shoulders, he looks like a typical avatar in the popular virtual world. But Edd is different.

His actions are animated not by a person at a keyboard but by a computer. Edd is a creation of artificial intelligence, or AI, by researchers at Rensselaer Polytechnic Institute, who endowed him with a limited ability to converse and reason. It turns out “Second Life” is more than a place where pixelated avatars chat, interact and fly about. It’s also a frontier in AI research because it’s a controllable environment where testing intelligent creations is easier.

“It’s a very inexpensive way to test out our technologies right now,” said Selmer Bringsjord, director of the Rensselaer Artificial Intelligence and Reasoning Laboratory.”

There’s more information in an article on Virtual World News. Apparently the goal is not to build interesting Second Life Bots using a variety of hacks, but to demonstrate human-like behaviour using more principled techniques.

“RPI is looking, initially, at a “theory of mind” for children, specifically with a false-belief test. In the real world, a child (age 4) would be shown a person placing a teddy bear in a cabinet. When the first person leaves, a second person would move the bear to another spot, like a refrigerator. When asked where the first person will look for the bear, they usually answer with the refrigerator due to a lack of understanding of other people. In Second Life, an automated theorem prover and procedures for converting conversational English into formal logic make up the brain of “Eddie,” the four-year-old avatar. When posed the above problem, Eddie responded as the human child would.”

Here’s a video of Eddie in action.

On Tuesday 22 January the agents mailing list (agents@cs.umbc.edu) will be offline between 21:00 and 23:00 UTC as we transition from Majordomo to GNU Mailman. Mail sent to the list at this time will bounce.

The agents list was begun in 1994 by Ray Johnson, then at the Lockheed Palo Alto AI Center and moved to UMBC in 1996. 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 manage the list and let users manage a wider range of their own subscription options. The list currently has about 2000 subscribers.

If you are a subscriber to either the UMBC agents or agents-digest lists, your subscription will be transferred to the new Mailman-supported list. Subscribers to the old agents-digest list will get a daily digest of messages. Using the agents administration page you can elect to receive messages as they are sent or to get them in digest form. We’ve assigned subscribers random passwords, so you will need to recover your password before making any changes.

You can edit your Mailman configuration now, but we won’t start sending out mail using Mailman until the Tuesday evening. I’ll send out an announcement via the re-hosted list when I know it’s enabled.

An address entered in the Mailman admin page must match your subscribed address exactly. If you are not sure which of your email address is subscribed, check the message headers to see if that reveals it. Failing that, you can try asking the old system by sending poor old majordomo@cs.umbc.edu an email message with the command “which ” in the message body, where is a string you believe to be in your subscribed address. As a last resort, ask me for help (finin@cs.umbc.edu).

You can continue to send mail to the list agents mailing list using the address agents at cs.umbc.edu. If the sending address is recognized as a subscriber, your message will distributed immediately and without moderation. Otherwise, you will be notified that your it awaits moderation, which might take a day or two.

In our old majordomo system, we maintained a separate list of additional pre-approved sending addresses. In general, if your sending address is not the same as your subscribed address, you should change the subscribed address. If you want to be able to send unmoderated messages from several accounts (e.g., your .edu and gmail accounts), you can always subscribe all of your accounts and disable email delivery for all but one.

Messages sent through the Mailman system will be available in an archive. The archive of old majordomo-era traffic is in disarray, but I think we have virtually all of the messages from 1994-2007. Eventually we’ll get it sorted out and online for posterity.

Our old moderation list was so inundated with spam and bounces from bad addresses that it became virtually impossible to moderate effectively. We anticipate that the new system will address both of these problems well and we will be thus be able to manage the moderation process better.

You can get more information about the list as well as manage subscriptions on the admin page and from the Mailman user guide. There are sure to be a few issues when we start using Mailman. If you have questions or suggestions about the list configuration, please let me know or send a message to the list if you think it should be of interest to the community.

Guaranteeing that you can take a hot shower is NP complete, at lest in one formalization the problem by Christina Matzke and Damien Challet in a recent paper.

Christina Matzke, Damien Challet, Taking a shower in Youth Hostels: risks and delights of heterogeneity, arXiv:0801.1573v1 , 10 January, 2008. … Tuning one’s shower in some hotels may turn into a challenging coordination game with imperfect information. The temperature sensitivity increases with the number of agents, making the problem possibly unlearnable. Because there is in practice a finite number of possible tap positions, identical agents are unlikely to reach even approximately their favorite water temperature. Heterogeneity allows some agents to reach much better temperatures, at the cost of higher risk.

Spotted on the physics arXiv blog.

xkcd has an IRC channel where its strange fans talk about even stranger things, some of the anyway. xkcd creator Randall Munroe discusses a common problem with IRC channels in a recent blog post ROBOT9000 and #xkcd-signal: Attacking Noise in Chat.

“When social communities grow past a certain point (Dunbar’s Number?), they start to suck. Be they sororities or IRC channels, there’s a point where they get big enough that nobody knows everybody anymore. The community becomes overwhelmed with noise from various small cliques and floods of obnoxious people and the signal-to-noise ratio eventually drops to near-zero — no signal, just noise. This has happened to every channel I’ve been on that started small and slowly got big.”

After laying out the standard approaches to controlling the problem (entry requirements, moderation, side channels) Randall describes a novel approach that fits oh so well with the xkcd community.

“And then I had an idea — what if you were only allowed to say sentences that had never been said before, ever? A bot with access to the full channel logs could kick you out when you repeated something that had already been said. There would be no “all your base are belong to us”, no “lol”, no “asl”, no “there are no girls on the internet”. No “I know rite”, no “hi everyone”, no “morning sucks.” Just thoughtful, full sentences.”

The idea’s implementation as a Perl bot sounds workable — when you violate the xkcd protocol by uttering a non-novel statement you are muted to prevent chatting for two second and the mute time quadruples for every subsequent violation. The bot forgives you after a while — your mute-time decays by half every six hours or so. You can read more about it on the xkcd blog or experience its tight rein on #xkcd-signal at irc.xkcd.com.

Not surprisingly, the channel is currently overwhelmed by chatters testing the bot to learn the finer points of its rules and how to subvert them. Hopefully, this is just a transient phenomenon and the robotic enforcement of novelty will evolve into something truly useful — a kindler, gentler moderator who can keep discussion from degenerating. But some serious tinkering will be required — common and repetitious utterances (”good morning”) are part of our social protocol, so this needs to be allowed to some degree.

FIPA is an IEEE Computer Society standards organization that promotes agent-based technology and the interoperability of its standards with other technologies. Jim Odell reports that FIPA’s P2P Nomadic Agent Working Group has released a draft of its specification. The group describes it’s focus as:

“The objective is to define a specification for P2P Nomadic Agents, capable of running on small or embedded devices, and to support distributed implementation of applications for consumer devices, cellular communications and robots, etc. over a pure P2P network. This specification will leverage presence and search mechanisms of underlying P2P infrastructures such as JXTA, Chord, Bluetooth, etc. In addition, this working group will propose the minimal required modifications of existing FIPA specifications to extend their reach to P2P Nomadic Agents. Potential application fields for P2P Nomadic Agents are healthcare, industry, offices, home, entertainment, transport/traffic.”

There is also a document from the Review of FIPA Specification Study Group that reviews and critiques the current inventory of 25 specifications.