What Has Your Agent Done for You Today?

Interview with Dr. Martin Griss — CoolTown m-pulse Online Magazine (September 2001)

Archival Interview: Originally published in Hewlett-Packard’s CoolTown m-pulse Online Magazine, September 2001. Preserved verbatim as a readable text archive (retired HP intranet navigation and interactive forms removed).

In an agent-filled future, users will trust their agents to tackle a range of daily problems and tasks. What these agents 'look' like, how they interact with each other and us, and what defines and limits their behavior are all questions HP Labs' Martin Griss is working to answer. By Amy Cowen

For Martin Griss, Ph.D. in Physics, Laboratory Scientist, and HP Labs' "Reuse Rabbi," the world of agents is less about holographic "butlers" that greet a Jetson-like family to report on the day's events and more about software components that can handle an assortment of tasks from ordering take-out to scheduling meetings. There's room for the friendly agent, but an agent with a visible avatar isn't necessarily a more effective - or more trustworthy - one, according to Griss.

The agents Griss works with may or may not have visible personas, but they will effortlessly shift roles and duties, adapt to current needs, make decisions based on context, and allow users to potentially pack even more into an average day. Looking at how personal assistants can function in multi-agent systems, Griss is particularly interested in how agents can and will work together, interacting with networks of agent systems as they perform their tasks.

How did you first get involved with agent development?

I do not remember exactly when I first became interested in agents. I have always thought of them as a kind of flexible software component, but my first real agent implementation activity started about 1994 when I and a university colleague worked on distributed instrumentation and robot controllers, using HP Vantara systems. Part of our work was to build a flexible visual programming kit - WAVE - and part was to build flexible components. We prototyped and demonstrated an agent-based process control system (curiously enough, in the area of electric power control in deregulated energy markets!), where lightweight agents would be located or sent close to the measurement sensors and transducers, while other agents would reside closer to information processing centers, or even the end-user.

Subsequently, we realized that a similar approach could be applied to systems and application management where a related but different form of management agent could be sent close to application processing elements and could filter management events, or even autonomously exercise some local control, before interacting with higher-level, more centralized, agents for more global control. Demonstrations and analysis of the system for OpenView led me to my current line of work in e-commerce agents and personal assistants.

When we talk about "agents," many of us imagine a "my butler" type of presence that appears on screen, interacts with us in familiar social ways, and can perform basic tasks to simplify processes. Shifting to agents designed to operate largely in a world of e-commerce - finding products, brokering deals, making payments, etc. - how "embodied" does the agent become?

In our work, we have largely focused on the functionality, customizability, and robustness of the personal assistant and supporting agents. We have done some work on providing a more natural interface using speech and natural language - for example, we can say "Please arrange a meeting with Reed and Craig next Friday." - and voice output, especially to voice mail, but the bulk of our work has been using simple web interfaces to create "one button meetings," instead of using three or four separate web interfaces and email exchange with consequent error-prone information exchange.

We have not explored the role of an anthropomorphic persona or avatar. There are other research groups who have concentrated most of their effort on natural interfaces, gestures, and animation to make the agent feel more flexible and more natural. It is, in fact, an interesting question as to whether these more natural interfaces increase our trust and comfort in using agents and in delegating increasing responsibility to them - reflect on whether the Microsoft agents "Bob," "Clippy," or "Einstein" inspire you with more or less confidence.

What kinds of agents have you experimented with for personal use?

The most recent agents I and several colleagues have built are focused on the personal space. The vision we have been pursuing is that of a "personal assistant" to a mobile person engaged in various kinds of personal or work activities. Imagine that I am walking around in a shopping mall, convention floor, or theme park, sitting in a theater, or traveling on business. I might be carrying a mobile phone or wireless connected Jornada, which might, in fact, have to be turned off. Traveling with me, and/or perhaps working back in the office or on my home server, is an active, intelligent, software component - my personal assistant - to which I can delegate several tasks that can be fully or partially handled for me.

For example, as I enter the mall or walk down the street in a shopping district, I might like my assistant to help me find stores which have items I'm looking for, perhaps negotiate discounts or special deals, find restaurants I might like and make reservations for me, or help me locate friends or colleagues who might also be in the same area. The assistant could actually make purchases for me, or assist in comparison shopping between items immediately available in a neighboring store, or available some other place in the region or on the web.

What are you working on currently?

Focusing on the personal and mobile aspects of agents, and for the moment downplaying the shopping and e-commerce aspects, we have been building a system we call "CoolAgent," which involves a group of personal, team, and service agents and assistants. These agents and assistants work together to manage calendars, meeting rooms, and teleconference lines to set up distributed meetings. While our demo is targeted at the mobile professional and the workplace, the same kind of personal assistant with supporting agents could help set up meetings and events between myself and my friends, my family, PTA, to better achieve work/life balance as my several roles interact.

Each personal assistant has access to a rich profile of personal information about myself, knowledge on how and when to access my calendar, what my preferences are for certain kinds of meetings, and how I like to be notified (by email, pager, voicemail, etc.) in different circumstances. In order to schedule meetings on my behalf, or respond to a meeting request from some other assistant, my assistant may be authorized to look into my calendar, to respond with my availability for certain meetings, and to reserve and confirm those meetings. At the same time, my assistant, and those of other participants, will negotiate with a meeting assistant, and perhaps a team assistant, to establish the best time, access the HP Labs' room reservation system to find an appropriate sized room with the right attributes and equipment, and also to reserve a teleconference line if some of the participants will be attending remotely.

Adaptability, autonomy, intelligence, and mobility are all features you ascribe to agents. Can you give an example of each in the context of a hypothetical agent named "Goop"?

Adaptability: Goop can adjust its behaviors to the changing environment, using downloaded extensions, parameter setting, rules, or machine learning. For example, if the user acquires a new device - the agent needs to "learn" how to filter information for it. If a new service becomes available, the agent might need to "load" a new protocol.

Intelligence: Goop can make powerful decisions and deduce some steps to relieve the user of specifying every step and decision in detail, and Goop can react to unforeseen (or relatively unforeseen) circumstances. Higher-level rules, policies, neural nets, planners, etc. are used to increase the agent's flexibility and level of abstraction.

Autonomy: Goop has some goals and an agenda to pursue (on its owner's behalf). It will proceed even if the user is not currently giving instructions and will try alternatives if other agents and data are not available. Intelligence helps increase autonomy. For example, a shopping assistant can be given a shopping list and budget, and it can buy directly on the user's behalf.

Mobility: Goop could do all of its work sitting on a central server, sending and receiving information from all over the web. Instead, a mobile Goop agent can have its agenda and knowledge packaged up and be sent out "in the field" to act on the user's behalf, gather and process complex data at the source, tour around from site to site, or launch itself onto a handheld or regional server to better support a nomadic user.

Another trait you identify is personality/sociability. As a developer, how do you approach this trait? How can personality best be integrated to end up with something that seems "natural" in the agent?

There is a lot of research and development going on, ranging from the construction of icons and images that feel natural through gesture and visual appeal to complex personal assistants using AI, fuzzy-logic, natural language, and speech I/O to communicate with humans as naturally as possible. Key elements are the ability to carry on a dialogue (perhaps spoken) rather than using a menu or browser, the ability to reason and make decisions without requiring rigid, fixed-format, complete input (using rule-based techniques, for example), and the ability to be highly customized to a user's preferences, priorities, and tasks (through rules, profile files, and machine learning).

The degree to which the system can communicate naturally, adapt itself to the user's style and needs, and appear to "understand," shapes the degree to which it shows a personality and acts socially with its user. In our work, we have just scratched the surface. We have used some simple spoken and typed natural language - "Please set up a meeting with Dick and Reed next Wednesday." - and are using a customization scheme using a combination of an XML-based profile that the user modifies and a set of Jess rules to represent and analyze preferences.

In one of your papers, you define software agents as "personalized, continuously running and semi-autonomous, driven by a set of beliefs, desires and intentions (BDI)." What kind of BDI parameters might underwrite the behavior and activities of a user's agent?

There are a number of models of agent-oriented programming. In general, when a set of components are very loosely coupled, communicate by exchanging high-level structured messages (such as KQML, or FIPA, or ACL based on speech act theory), and exhibit significant autonomy (having goals, pursuing them independently, and being able to find and request assistance and services from other such components), we call them agents.

It's convenient to use an anthropomorphic abstraction when decomposing a system into agents and when assigning responsibilities and programming an agent. We tend to ask questions like "What does the agent know? What does it want to do next? Can an agent know what another agent believes?" In some cases, this has been formalized into the "beliefs, desires and intentions (BDI)" model, which can be used to reason about what other agents know and might do in certain situations; in other cases, the use is more informal. In our work, we are much more informal since we do not believe we can have total control over what agents written by different programmers will do. So instead of thinking about BDI in regard to other agents, we use knowledge, goals, and rules to program our agents and have them exhibit specific behaviors and follow prescribed patterns of message exchange (message protocols), which involve specific use of words (the ontology).

How are BDI and artificial intelligence related?

First of all, we want our agents to be flexible and autonomous. That means they must be able to handle complex alternatives, ideally be able to handle situations that have not been fully detailed by the programmer, and be able to adjust to changing situations and priorities. So while all of this capability can be programmed directly in a procedural language such as Java, it is much more convenient to use toolkits that allow us to program at much higher levels of abstraction. These include a number of AI techniques: rule-based and knowledge-based software (using a rule interpreter such as Prolog, CLIPS, or Jess), planners and goal-directed techniques (so that a goal can be stated at a higher level and decomposed into achievable sub-goals), and fuzzy logic and machine learning techniques (such as neural nets and Bayesian reasoning). Not all agents need to use all or any of these techniques, but they make it easier to give the "illusion" of autonomy and intelligence. The challenge in using such techniques is making it easy for the user to understand what will happen under different situations.

You've written about the flexibility (and desirability) of a "dynamic agent" that can "adjust its capabilities and play different roles to accommodate changes in the environment and requirements." How do dynamic agents operate differently than their more static counterparts?

We must design agents assuming that other agents may or may not be there, and may or may not respond as expected - at any time. An agent is free to respond with "not understood" or " failure" or "refuse." Our agent has to be willing to deal with these [responses] and other timeout or error conditions. At any time, other agents may be added into the system, able to publicize their ability to respond to requests of certain kinds. Part of the approach is to make it possible for agents to have a rich repertoire of behaviors and the ability to dynamically load or learn new behaviors.

Agent construction seems particularly tricky given standardization and interoperability issues. What steps need to be taken industry-wide to support a widespread system of interoperability that can support the construction of a robust network of software agents?

Even in the absence of agents, we have the ability to create new standards, new rules, new business models, new roles, and common jargon and language for communicating. However, many of these have been represented informally, or at least in an ad hoc machine-inaccessible form (e.g. consider the tax code or legal system; EDI is more formal but 1-to-1 ad hoc). If we want autonomous intelligent agents and web services to be able to dynamically discover each other, and establish communication and interaction, and then begin to negotiate, the standards and rules have to be represented in a more formal structure. This is where work on ontologies, agent standards (like FIPA), and web service standards (XML, SOAP, UDDI, ebXML, RosettaNet, semantic web) will play a significant part. These communication and representation standards are now beginning to emerge. Some parts will become ubiquitous, while others will be associated with particular e-systems or ecosystems; perhaps this will be an opportunity for translator agents and services.

The variety of agent systems that exist - Aglets, Grasshopper, Jackal, Voyager, Zeus, etc. - hint at an already robust agent-oriented landscape, but can all these different agents really communicate with one another and work together on a variety of planes? Or is the agent owner inevitably going to end up with an agent that can only talk to "some" agents but not all of them because they use some "other" system (similar to OS-issues)?

Indeed this is a problem - I have found over 100 different agent systems, many with their own agent-modeling communication languages. However, recently FIPA has established a set of standards, which many of the newer agent systems try to follow. In particular, there is a standard for an agent communication language (ACL) and the model of a standard agent platform and agent services, which is intended to make it easier for separate agent systems to communicate. Also, recently a number of agents systems (such as Jade, Zeus, FIPA-OS, Grasshopper) have started to release Java-based open source, or at least "source-available," platforms. These kinds of standards are increasing the possibility and probability of interoperation.

Finally, a large-scale, multinational interoperation effort, called AgentCities is expected to act as a forcing function for the development and testing of ontologies, agent communication languages, agent platform, and agent services. The idea of AgentCities is for several separate groups to establish an agent-based service within/for a number of different cities such as Paris, Tokyo, San Francisco, or perhaps associated with the Olympics, to make available services such as restaurant booker, shopping assistant, event guide, etc. HP Labs is playing a key role in helping define AgentCities.

In "Games at Work Agent-Mediated E-Commerce Simulation," you talk about using games to analyze and test e-commerce strategies and, ultimately, to accelerate development while reducing errors, scalability, and integration issues. How complicated - and costly - is this kind of game construction? Is it feasible for big e-commerce sites to move to this kind of testing and analysis prior to beginning full-scale development?

I have to admit that this vision has so far been only partially realized in our work. When we started, the original idea of the game scenario was to provide us a protected "sandbox" in which we could develop and evaluate agent and web service technology. Using game theory and agent-based game participants to test the particular slice of the shopping or the commerce environment is definitely feasible and might help businesses get a better handle on their pricing, selling, negotiating and advertising models.

The "CoolTown Impatient Shopper Scenario" is an interesting envisioning of the way agents might take shape and how their presence could alter ordinary/daily e-commerce-related tasks. As agents begin to become a part of daily life, how do individuals avoid being bombarded - or exhausted! - by the information the agent can make available?

With agents available to handle many of the tasks that once required human involvement, how do we safeguard the human factor? After all, the user need never leave his/her house if a good agent is on scene.

In part, the question is based on a false premise - it will be a long time before a truly intelligent, general-purpose agent that can do everything for us will be widely available (if ever). It is also based on a possibly false assumption that people would prefer to stay home and interact only though the web and their agents. People have expressed these concerns before about the phone, TV, and home PCs. Even today, we are seeing two modes - there are those who stay home and surf the web and live in chat rooms, rather than meeting face to face - but they do communicate more than before, just as teenagers would and still do hog the phone. But other communities of young people use personal communicators and phones to arrange meetings in malls, communicate with each other across tables, in classrooms, etc.

You can see these two approaches in two science fiction books I like. In Snowcrash (as in Matrix and Netforce ), people almost always only interact through the system and take on virtual personalities to overcome the tedium of their real lives. In the Asimov I Robot stories, on the other hand, the agents (aka robots) have a physical embodiment and are more like servants who go with the individual and help them do more things than they would without the help. In both approaches, there are cautionary tales of overdoing the dependence on the primary mode. In some cases people "disappear into the system"; in others, the robots so vastly outnumber the people, that it's really about a robot society, not a human one... so, no good answers yet.

How far are we from having an agent in every home?

It depends on what we want the agent to do, and whether it comes bundled with a mobile phone, lives as a service in our ISP, telco, or cable provider, or is set up by the individual. Perhaps the agent is not "in every home" - instead, it lives in a secure place on the web and delivers services to us through the web or phone network. Better to say "when every person has their own personal assistant." Already, some buying agents and comparison-shopper agents are accessible to many people (e.g., eBay) over the web.

It is fairly easy to imagine some kind of agent-based calendar assistant or email assistant very soon for those who use PCs, Jornadas or Palms. It is also easy to see simple agents appearing in cable set-top boxes, just as the VCR is being displaced by "intelligent" video recorders with learned preferences.

It's a bit further off if we want agents to support buying from phones, or more intelligent interaction through speech, or more collaboration between the personal assistants of multiple people... then each person in the interacting community needs to have one, they always need to be running, they need to be secure, etc.