|
Presence Examples From
APC (Australian Personal Computer, "Australia's premier computer magazine") AI think, therefore I am APC November 2003 By David Braue Virtual agents feature Computerised characters that look, sound, move and seemingly think like real people are emerging from the realms of science fiction into everyday life. David Braue reports. Making computers human is an idea as old as computers themselves, and what was initially a wild science fiction fantasy is gradually turning into fact. From the chilling 2001: A Space Odyssey’s HAL 9000 to robotic newsreader Ananova and Jar Jar Binks, virtual creatures have become part of our collective culture. Much more than entertainment is at stake, of course. The potential of computerised agents or entities that are autonomous, self-directed, reactive and social — just like humans — can be estimated only in the realm of the imagination. Already, such agents have been built to present the weather on mobile phones, drive trucks, monitor environments designed to support life on other planets and perform many other sophisticated tasks. Computers are good at doing what they’re told, but in this field they’re required to reach their own conclusions. The complex computer code beneath their “skins” is designed to make them react to situations like real people do — unpredictably. Just how far we have come was evident in Melbourne earlier this year when more than 450 researchers from 29 countries attended the second annual Autonomous Agents and Multi-Agent Systems conference. The purpose was to share work in a steadily growing field combining the personality theory of psychology with the latest developments in artificial reasoning and unstructured data analysis. Someone to talk to One of the first agents to command attention was Eliza, a simple yet intriguing virtual psychoanalyst who hung out her shingle in 1966. Developed as an experiment in natural language processing, she has been frustrating users ever since. In 1950, Alan Turing made the postulation (in what would become known as the Turing Test of Artificial Intelligence) that a computer could be said to be thinking if we couldn’t tell the difference in reponses to an enquiry between that of a computer and that of a human. Joseph Weizenbaum, a scientist at the Massachusetts Institute of Technology, put the test to task in 1966 by creating Eliza. For its time, Eliza was a sophisticated text-parsing engine that simulated a psychoanalysis session by answering input with related questions. The title of Weizenbaum’s paper revealed his motivations: Eliza: A Computer Program For the Study of Natural Language Communication Between Man and Machine. In the last two decades, Artificial Intelligence (AI) algorithms have matured to create far more sophisticated agents, and researchers can now use advanced graphics technologies to create convincing characters in the form of 2D cartoons or 3D people. At the University of Southern California’s Information Sciences Institute (ISI), a breeding ground for international AI research, the latest agents are being imbued with a range of human characteristics: guilt, anger, defensiveness, and so on. Personalities that determine reactions to situations can be defined using onscreen sliders. ISI has been collaborating with the US Army on a virtual reality project to populate a fictional war-torn village with agents, each having its own personality. Trainee peacekeepers interact with volatile onscreen characters on a massive projection screen. Get frustrated with a mother whose child has been injured in a car crash involving a Humvee, and she’s likely to become hostile. The high-end application runs on refrigerator-sized SGI computers in a lab at ISI, but the technology has also been scaled down. One program, designed to help mothers of children with cancer, uses a mother’s alter ego named Carmen. Carmen runs on a standard PC and has been successfully trialled as a potential way to reduce hospital social workers’ workloads. This sort of interactive agent could ultimately improve even general learning by providing omnipresent mentors to watch students’ learning processes and automatically offer relevant advice, says Dr W. Lewis Johnson, director of ISI’s Centre for Advanced Research and Technology in Education. The challenge is finding the right balance between constructively interrupting users and annoying them. “We’ve been looking at the way people praise, the comments made to other people based on the nature of the social relationship between them, and the degree of trust that’s been established. Then we try to emulate that in the agents,” Johnson says. “For example, it might start with more polite interaction and gradually transition to a more direct style once the computer and student have developed a collaboration.” Productive mentor-student relationships take weeks to develop in a conventional classroom and can be easily broken if the teacher makes mistakes or is otherwise fallible. In a computerised setting, the challenge of building trust becomes difficult if the agent is unconvincing. Stuttering jerks One of the biggest problems is that many agents just don’t look real. Like the stuttering, jerking Max Headroom, they have poor speech-to-lip synchronisation, awkward movements, unconvincing features and unnatural expressions. The relationship and effective transmission of information is disrupted because the human participant becomes distracted. One French study evaluated full-body 2D agents presenting different types of information. Respondents awarded higher scores to an agent programmed with a larger smile and disliked a character whose eyes were barely visible behind his glasses. Others were unimpressed with an agent who impatiently folded his arms. A similar study in the US measured the trust put in agents based on their appearance after research suggested that 55% of the meaning in human-to-human contact came from facial cues. It was found that the perception of agents was affected by smiling, random movements such as blinking and eyebrow changes, eye gazes and bodily gestures. In other words, the importance of appearance when two flesh- and-blood humans communicate isn’t diminished when one of the participants is replaced digitally. Creating convincing facial behaviour increases perceptions of trustworthiness. Unconvincing body gestures have the opposite effect. As a result, researchers suggest that agents be presented onscreen without bodies. University of Tokyo researchers have taken this approach in using Multimodal Presentation Markup Language (MPML) to display an animated agent that reads news on mobile phone screens. Even when an agent’s appearance is impeccable, its behaviour must be carefully controlled. Over-attentive or inaccurate agents quickly become a nuisance — as millions of users of Microsoft’s Office suite have known for years. Ah, Clippy, the animated paperclip, and his host of friends, who constantly watch what the user is doing and offer advice when they feel it’s appropriate. Microsoft took months to train them, working with hundreds of beta customers with a special “instrumented” version of Office that recorded each function they used. The data was then collated to find out which functions were used most frequently or rarely, in which combinations, and in what types of documents. “It became obvious that users anthropomorphised the computer,” says Chris Pratley, group program manager for Office authoring at Microsoft’s Redmond headquarters, referring to the tendency for people to give non-human objects human characteristics, from Donald Duck to the plastic boxes on their desks. “They would feel sorry for themselves if they thought the computer had done something wrong, and feel bad if they did something to the computer. Research suggested that if we could somehow make the computer have more of a social interface, users would relate to it with more of the interaction they would have with a human being.” Inevitably, Clippy’s charm wore off and advanced users quickly tired of his interruptions. While acknowledging that many users turn off the agents as a matter of course, Pratley says they are still useful as a familiar delivery mechanism for the program’s many tips and for feedback. The agents remain in Office 2003, and the help system they access has been extended online to a broader range of tips that are constantly being scored by users. Microsoft’s agents have shifted from being interfering nuisances to more passive conduits to information. Mechanical assistants Focusing more on the mechanics of artificial behaviour, disembodied agents are being used in a wide range of applications as a means of collecting new information rather than regurgitating it. Agents are all over the Internet, across which search engine “spiders” interactively locate and index sites, and are also common in subscription news services. Rather than having personality-based controls, bodiless online agents are directed by parameters such as where, how far and how deep they should search for information. Many researchers believe such agents will become pervasive personal assistants, helping people keep up with a constant flood of information by proactively sorting, cataloguing and presenting it in a meaningful way. Liz Sonenberg, a human-computer interaction researcher at Melbourne University who helped bring AAMAS 2003 to Australia, sees agents as crucial for filtering large volumes of data for presentation on the limited screens of PDAs. Her work revolves around teaching computerised agents to figure out what users are doing and load the information they are going to need in advance. Agents could also work proactively on the user’s behalf. For example, if a PDA equipped with a global positioning system knew its user was running late to a meeting (a conclusion reached because the person wasn’t in the right place at the right time), it might automatically contact the other party’s PDA to negotiate a new time. “If you have limited space but access to a lot of data behind the device, you should be presenting it to users so they can get what they need in a very understandable form,” says Sonenberg. Agents are also proving useful in far more sophisticated applications at organisations such as the Bureau of Meteorology (BoM), which must sift through mountains of data collected from around the world, 24 hours a day. This data is normally analysed using an NEC supercomputer, which updates forecasts every three hours. But, the global focus means meteorologists may miss local weather changes. The BoM uses agents to fix the problem, teaching them to watch for even small changes that human workers may miss. The agents were built using JACK, a toolkit from Melbourne’s Agent-Oriented Software. The software creates Java-based agents according to defined behaviour rules, and the agents can raise the alarm when, for example, atmospheric pressure has changed by more than 2 millibars, signalling an approaching change. JACK’s flexibility has made it popular with organisations keen to experiment with agents. Airservices Australia is interested in agents that automatically detect conflicts between military and civilian flight plans, then help controllers find alternative routes. The Defence Science and Technology Organisation is using it to help test combat planes. Even the UK’s Ministry of Defence is using JACK to build agent-based virtual soldiers that feel fear, fatigue, distraction, anger and so on. Military planners hope these “soldiers” will offer insight into how real people react in difficult situations, identifying areas where individual personalities may override training. Agents can be used to help people — or replace them — in all sorts of applications, and it won’t be long before they are indispensable everyday companions. Behind the wheel In the film Maximum Overdrive, possessed automobiles and trucks make it their mission to exterminate humans with extreme prejudice. The result is an extremely silly movie filled with long shots of driverless cars making absolute menaces of themselves. But at several mining sites around Australia, where hulking robots drive mining trucks with cool and calm, this movie doesn’t seem so ridiculous. However, their driving skills don’t come from some inherent machine-to-machine emotional connection. Rather, they’re the result of extremely smart programming work by a team of over 100 researchers at Sydney University’s ARC Centre of Excellence in Autonomous Systems (CEAS). “We have agents embedded in trucks, excavators and individuals [robots] in order to mine the right material at the right time,” says Hugh Durrant-Whyte, research director at CEAS. “We do not approach it at all from a human point of view — robots are really physical embodiments of agents. They won’t discuss Plato with you, but they can work 24 hours a day and have cooperation and negotiation strategies [to interact with each other].” Working with commercial partners, CEAS’s goal is to apply intelligent agent technology to real commercial applications — particularly those involving regular, predictable actions in a defined space. That’s why its first partner was stevedoring giant Patricks, which worked with the group to apply agent intelligence to loading and unloading containers at the docks. In a simulated micro-economy, each container gets an agent of its own, which bids against other agents to see which container is willing to pay the highest “price” to be unloaded before others. Safety considerations and the complexity of moving heavy trucks through changeable environments made mining automation the second major area of work for the group. Still more complex is firefighting, where specialised agent “experts” would help humans by monitoring firefighting assets, predicting the best strategy and analysing satellite imagery to predict a blaze’s spread. “Agents are a very good way of breaking down a problem into solvable sub-units,” says Durrant-Whyte. “Agents tend to be focused on a particular domain, but they’re able to take into account very unpredictable environments and how things change. You wouldn’t want to say they behave like humans, but they’re much more flexible and able to choose appropriately than old-style software.” Controlling virtual agents Computerised agents can consist of two main layers: the presentation layer, which projects the face and other graphical elements for creating a person or character; and the logic layer that encapsulates the AI algorithms and decision-making. Many agents have no presentation layer, working invisibly in the background, but the logic layer is common to all. Logic-only agents are generally autonomous entities that can accept comparatively straightforward parameters, such as the text they should be searching for. But agents with a presentation layer are much more challenging for researchers to create, involving appearance and human-style interaction. Just drawing the faces takes a lot of work, which is why many researchers adopt widely available packages for drawing and animating agents onscreen. The popular Microsoft Agents loads Web browser plug-ins that allow Web pages to be illustrated using a variety of characters. Other options are EptaMedia’s EptaPlayer, with plug-ins to play animations created using the companion EptaPublisher, and Digital FaceWorks. Showing character is only one part of an agent. Most also rely on a text-to-speech (TTS) engine such as Lernout & Hauspie’s TTS3000, complemented by a Speech Application Programming Interface (SAPI) 4.0 package, with optional use of voice recognition through IBM ViaVoice, the Microsoft Speech Recognition Engine, or similar technologies (all three are available from the Microsoft Agent download page). Another TTS engine is the University of Edinburgh’s Festival Speech Synthesis System. Once the components are installed, a way to drive them is needed. Specific applications may call the agent or TTS APIs directly, while other approaches control agents by seeding Web pages with tags in languages such as Virtual Human Markup Language. VHML controls agents using standard tags such as <smile>, <confused>, <sad>, <agree>, <concentrate> and so on. VHML includes sub-standards such as Speech Markup Language (SML), which allows Web pages to drive TTS engines; Facial Animation Markup Language (FAML); and Body Animation Markup Language (BAML). Perfecting the pilots The Defence Science and Technology Organisation at Fishermens Bend, Victoria, has established itself as a hotbed of aeronautical innovation. The black box flight recorder was invented at the site, which spreads across a series of buildings that were once aircraft hangars. So too were the sonar-silent tiles for Collins Class submarines, and the Jindivik radio-controlled target drone sold around the world. The site is also where Defence tests pilot-plane interaction, with life-size, working cockpits that allow engineers to try various configurations, and conduct psychomotor tests on pilots’ reactions and attention during simulated battles. New plane designs are also put through their paces in detailed virtual simulations. The DSTO is using agents built on Agent-Oriented Software’s JACK Intelligent Agents platform, working closely with AOS to improve its design. The DSTO’s focus on computational simulation makes it interested in agents for their ability to repeat tasks with great precision rather than the decision- making and personality-based qualities wanted for many commercial and research applications. When testing new equipment or measuring the subtle results of small configuration changes, repeatability is paramount — and human pilots always do things a little differently no matter how hard they try not to. “This represents the separation of the decision-maker from the physics,” says Dr Simon Goss, the DSTO’s head of operations research capability. “Before, we tended to use mathematics to model the airplane, then the pilot. But in the testing phase, we don’t want agents to be superhuman. We want repeatability and traceability, so we can say that the agent did this for this reason. We want them to do routine things in a sensible manner that matches military doctrine.” In recent experiments evaluating a new type of plane with an extended sensor suite, the DSTO air analysis team used agents to simulate the plane’s behaviour and assessed the ability of human pilots to job-share with agent-enabled simulation environments. The agent’s role as virtual test pilots was reinforced when the DSTO put them into the real-life application SWARMM, which was developed in the mid-1990s, to simulate the dynamics of air missions and pilot reasoning. They have since added realism to broader combat simulations in which multiple agents interact with each other and with human commanders.
|