The "personal agents" are programs or robots that will know how to perform boring monotonous actions in our place, which contemporary technology not only does not free us from - but even increases their number
Israel Benjamin, Galileo
When we are glued to the computer screen while following the news, the stock market or an auction; Every time we receive the same type of email for the thousandth time, and perform the same action in response; Or when we try again and again to reach someone on the phone or are transferred to an endless record (or worse, are forced to listen to self-praise about the quality of service of the organization we are trying to reach), the thought creeps in: "There must be a better way!". After all, technology was supposed to save us repetitive actions, not add to them...
A possible answer to this requirement is a "personal agent", who will receive instructions from us and act autonomously to carry them out. Already today there are many such "agents" for specific tasks. For example, in the field of online auctions, there are software programs that allow a person to wait until the last minute before the auction closes and then raise the offered price up to the limit set by the buyer for that item. If part of my work routine is logging into several websites to check if anything new has been posted on them, I can transfer the task to software that will notify me of any changes in content - for example, by sending an e-mail. More on the subject
Such programs are the first generation of autonomous agents. Although it is difficult to regard them as intelligence, they herald the next wave of software that can truly represent us, will act flexibly to achieve goals without requiring us to define exactly the course of action, will negotiate with other people (or, of course, with their agents...) and will finally set us free - The end of the fandom of technology.
What is an agent?
The idea of creating agents has been proposed many times by computer scientists working in the fields of artificial intelligence, artificial life, natural language understanding, and more. It is not surprising that there are many definitions for this concept, and some of them even contradict each other. Despite this, you can find some features that are common to the different settings:
The agent operates within a dynamic environment, receives information from it about the changes in the environment, processes it and can take actions that affect the environment. In other words, the agent is equipped with sensors, effectors (ability to manipulate the environment), and some process that links them. It is not difficult to meet this requirement. Although it is also suitable for a very complex entity - for example, a person - it also exists in a simple thermostat, which consists of a temperature sensor, a circuit breaker that controls the operation of the oven, and a very simple control process: when the temperature is high enough, the thermostat turns off the oven, and when it drops, the oven will be turned on again. Indeed, the thermostat frees us from continuously monitoring the temperature and turning on the circuit breaker ourselves.
The agent has a goal that he works to achieve. The thermostat works to keep the temperature within a certain range. The auctioneer works to purchase the item at the lowest possible price.
The agent has autonomous and persistent behavior. Unlike a car or word processing software, for example, it operates continuously and does not require a person to dictate its actions to it. "Smarter" agents have an additional meaning to their continued existence - a feature called "persistence": they also have memory, which preserves past data and thus allows them to learn and improve their performance in the future. Such an agent is adaptive - it adjusts itself and adapts to the environment.
Multiple agents can communicate. In simple cases, the agent can inform the person he represents of his progress in achieving the goal. When the agent is a computer program, messages are usually sent to e-mail or SMS. As we will see later, the situation is more interesting when agents talk to each other.
More advanced agents are endowed with flexibility. A flexible agent is not limited to rigid decision processes. In auctions, a rigid process might be "increase the price in dollars above the last bid up to an upper limit of X dollars". A flexible agent will employ more complex tactics, such as a high initial bid when the price is still low (to perhaps make some competitors conclude that they have no chance), or even searching for identical items in other auctions and deciding on the bidding policy according to the number of possibilities to obtain the same item and according to the development of sales the others.
The mobile agent
When the agent is not software, but a machine whose environment is the real world - a robot - it is often endowed with mobility: the ability to move from place to place to gather information or perform its tasks. Even in the virtual world of computer programs, portability has meaning: given suitable conditions, programs may copy themselves from one computer to another. A well-known example of this is a computer virus (or worm), which spreads itself over computer networks to achieve the goals of the software author - advertising, causing damage or perhaps gathering information. Portable software is not necessarily harmful: in recent years, there has been a great deal of interest in creating software that can find the most suitable computer for itself at each stage of performing its tasks - to, for example, switch from a busy computer to a free computer, or reach a computer with specific capabilities needed at that moment.
We will examine one way to create agent software, which will help automate one of the most time-consuming tasks for information workers: reading e-mail, sorting it into groups and starting to process it. Let's ignore "spam" for a moment, and assume that all messages are legitimate and we are interested in receiving them. However, we would like to delve into only a small number of the messages. All other messages fall into one or another category of messages that require simple and quick handling. For example, let's assume that Orly, a production manager, usually deletes messages distributed to all employees about offers for shows; She files weekly reports in a report folder; And when a request for the purchase of raw materials arrives in her mailbox, she opens the resource management and procurement software, and types in the required material code to view the current inventory level and the planned production processes that will need this material.
Many e-mail programs allow you to define "rules" that enable a certain level of such automatic management. For example, Orly can create the rule: "If the sender is 'Human Resources' and the subject contains the words 'Show schedule,' delete the message." The problem begins when a large number of such rules must be created, and maintained so that they operate in a changing reality - for example, after the human resources department starts sending such messages with the subject "shows and entertainment". Most of us would prefer to continue the boring routine of individual handling of each mail item, if the alternative is to analyze each error, identify the change that caused the error, add or change rules, and check the correctness of the change we made.
If we could operate an artificial intelligence approaching human intelligence, we could treat it as an assistant or secretary guides; However, we are still very far from such an ability. We cannot create software that will read the message, understand its content, ask us for guidance on handling such content, and be able to understand how to generalize - and when not to generalize - this guidance to subsequent messages. These problems are not prevented by programs that exist today to identify and block "spam", for example by using probabilistic inference networks for the purpose of learning (see "The priest and the probabilistic intelligence").
From here to the automatic inference of ways to classify the e-mail and respond to it, the path is short, at least on the principle level: the agent "watches" Orly's actions, and keeps in his memory the actions she performed and the content of the messages. After enough information has been accumulated, the agent tries to identify patterns - in fact, it creates and maintains on its own rules similar to those we described above. Indeed, several such projects have been developed or are under development, in academic research institutions and commercial companies. Recently, a new use was found for such software, when a large company is suspected of criminal offenses, and the police seize a large amount of e-mail written or received by the company's employees in order to check these suspicions, but have difficulty finding the incriminating information in the mountains of mail. Some recent academic studies compare the success of automatic classification methods by running them on about 500,000 messages, published by the American commission that investigated the Enron scandal.
Sociology of agents
If one agent can be useful, what can be achieved by allowing multiple agents to work together? For example, what would happen if we allowed Orly's email management agent to share information with similar agents serving other employees in the company? On the one hand, Orly's preferences are not shared by all employees: other employees may be interested in information about offers for shows, for example. On the other hand, employees who do not know how to handle procurement requests will be able to use the knowledge gained by Orly's agent.
One way to reconcile these possibilities is to decide that when an agent has learned clear preferences of its human operator, it will ignore other people's preferences. On the other hand, when a message appears for which no clear preference has been observed, the agent will check what are the typical actions of other employees and suggest them to the user. For example, a message about discounted tickets to a show will be automatically deleted at Orly, but with a new employee, the agent will display a menu next to the message, with the options learned from other employees, such as "delete", "file in the show folder", or "open the ticket ordering software". It is understood that such an idea reveals to some extent the preferences of employees, and thus may harm their privacy, but this important issue goes beyond the scope of the current column.
Collaboration
It is almost certain that the social behavior of agents will be fundamentally different from that of humans. Advantages may also grow out of the technological limitations. Let's take for example the example of agents trying to get the best price at auctions. Suppose that a hundred products of a certain type are offered for sale, at a starting price of one hundred shekels. The product is popular, and thousands of consumers are interested in it. For the sake of simplicity, let's assume that the consumers are all the same in terms of their financial resources and the value they attribute to the product. In the normal conduct of an auction, the multitude of interested parties will lead to competition and the increase of the price up to the maximum that the consumers are willing to pay, and since they are the same, the winners will be determined arbitrarily - for example, according to the timing of the submission of bids.
On the other hand, if the thousand agents cooperate, they will be able to choose in a fair lottery which of the hundred will approach the tender, and each of them will pay one hundred shekels (similar solutions exist even when the competitors differ in the maximum amount they are willing to pay). Such cooperation is not possible between humans: some losers will violate the unwritten agreement and submit higher offers. In this case, the advantage of the agents is precisely in simplicity and transparency: the scenario presented here is only possible if we are sure that we understand the behavior of the agents, and know that they do not have enough intelligence to consider violating agreements. As is well known in game theory, even this scenario is unstable: the appearance of one user is enough to change the behavior of his agent, or even the suspicion that such a user may appear, to renew the price war.
The following answer has already been offered to such problems: if the "honest" agents unite, identify the violators of the agreements and refuse to cooperate with them, success is still possible. In any case, the purpose of this example is to illustrate the possibility of agents not only to speed up and facilitate existing processes, but also to bring about far-reaching changes. We note that the auction operators will not welcome such collaborations, and that they may change the rules of the game, or employ their own agents, which will create an even more interesting and dynamic environment than the one described so far.
How far can agents develop independently? An ambitious research project called NEW-TIES examines these questions. Researchers from five European universities (two from England, two from the Netherlands and one from Hungary) will develop a software environment in the coming years in which about a thousand agents will move and communicate with each other, on fifty computers provided by the universities. In the most conservative formulation, the goal of the project is to learn how to create agents that adapt themselves to an environment whose properties are not known in advance (a more ambitious goal: "to create an artificial society that will develop mental and linguistic abilities on its own, and try to understand how the world in which it operates was created").
One of the initial requirements is to make the agents learn how to find "food" for themselves by growing it or trading with other agents, finding mates for the purpose of creating "offspring" agents that receive some of the "parents" traits, and more. The agents will not be equipped with a predetermined language for communication between them. The researchers hope that the agents will discover ways to cooperate, and will develop a new language of their own for this purpose. If this happens, it will be a very significant breakthrough in artificial intelligence research, which will contribute not only to our ability to recruit smart and trained assistants, but also to our understanding of the processes that make us, humans, what we are - creatures defined by the society in which they exist, in the process that they shape the face of that company.
