Lifelike Autonomous Agent (Laa – Local, Autonomous Systems)
Pattie Maes, Head of the Autonomous Agents Group on the Media Lab of the with, is a pioneer in the field of agent research. One of the central applications of agents provides you in the entertainment area. Here is the design of the interaction with people especially important and groundbreaking for agent research.
Pattie Maes is Associate Professor at the Media Lab of the with which you are the autonomous agents group reason that you are heading now.
The relevant new area of artificial life (KL) tries to explore biological life and to understand by the production of artificial life forms. Around Chris Langton, the burden of this area, to paraphrase, is the goal of artificial life, to renew the life as it could be to understand the life as we know it. Artificial life is a very gross science area, which adds different topics such as artificial evolution, artificial oksystems, artificial morphogenesis, molecular revolution and many others.
The KL research is interested as well as Ki research at a synthesis of continuous autonomous agent (AA). AA are computer programs that inhabit a complex, dynamic environment in this environment autonomously and orient themselves and thereby realize a number of goals and tasks, for which they have been constructed. The project to build an AA is as old as the field of artificial intelligence itself. The KL community has developed a completely different approach for this goal, which is more focused on fast, reactive behavior and on past and learning than knowledge and reason. This approach is inspired for the coarse part of biology, and especially of ethology trying to understand the mechanisms that use animals to realize customized and successful behavior.
AA can take the most diverse forms according to the type of environment that you inhabit. If the environment is the real physical environment, the AA ames the form of an autonomous robot. Alternatively, one can build 2D and 3D-animated agents, which inhabit simulated physical environmental environments . Ultimately, so-called knowboties, software or interface agents are corporate creatures that inhabit the digital world of computers and computer networks. For all these types of AA there are clear applications. Autonomous robot Z.B. have been developed for surveillance, exploration and other tasks in environmental, which are inaccessible or dangerous for people. There is a long tradition in the production of simulated agents for training purposes. Interface agents are also recently considered a mechanism that helps computer users to cope with the work and information relief.
A possible field of application of agent research, which has been given a surprisingly little attention, is the entertainment area, which will increase significantly in the coming years, as the traditionally rich money source of agent research, the defense industry has been shut down. The entertainment industry is a strong industry, from which one expects that he expands in the next few years. Many forms of entertainment use figures acting in an environment, z. B. Video games, simulated horse riding, films, cartoons, animatronics, theater, doll games, specific toys and even party lines. Each of these forms of entertainment could benefit from the use of autonomous, semi-intelligent agents as entertaining actors. The entertainment industry is an enthusiastically oriented and very challenging scope that will expand the limits of agent research.
The challenge of modeling entertaining game figures
Various forms of commercial entertainment are presenting at present automated game figures. Most are simply simple: they show a very predictable behavior and do not very convincing. This is especially in such figures the case with which one can interact in real time, z.B. for video games. If automated characters show intelligent behavior, it is usually completely mechanical and not interactive and usually the result of a casual and labor-consuming production. An example of this is the behavior of the dinosaurs in the movie Jurassic Park.
In recent years, however, there were a few exceptions. Some researchers have made animated films with the agent technology. Instead of presenting accurate movements instead of the animated figures, they are modeled as agents that exports actions in response to the environment they perceived. Reynolds modeled bird and fish warms by specifying the behavior of individual animals that formed the swarm. The same algorithms were used to determine certain behaviors of the bat in the film Batmann II to create. Terzopolous modeled extremely realistic behaviors of fish that are couples and eating, learning and hunting. These models were used to make short, entertaining trick films . In addition to the caught work, some researchers have used agent models to build interactive animation systems in real time. the Female Welt of Bates allows the user to interact with a world of living beings, which are called boggy. In this pioneering work, a user interacts with the world and their creatures by controlling the movements and behavior of the boss directly with a mouse and the keyboard. The weighing have different inner needs and a coarse bandwidth of flying, which leads to pretty complex interactions.
Realistic fish behavior in the simulation of Terzopoulos
Fisher’s Menagerie allows the user to interact with the animated agents in real time by putting on a cyberspace helmet. The agents of the Menagerie normally drove a single complicated behavior such as Z.B. the formation of a herd or a swarm. Tosa put on neuronal networks for an artificial baby, which reacts to a feel-fashioned way on lute, which makes a user looking into the cradle. The Alive system, which will be described below, allows the user to enter the user to enter a virtual world and to use all its body to interact with animated autonomous agents.
The Alive System
In addition to the computer-animated agents, there is the remarkably ingenious project Julia, an autonomous agent who lives in a text-based MUSE system (Multi-User Simulation Environment). Julia can move in the environment and this mapping, talking to the players, talking to entertainments and talk about telling them, support messages between the players and help them with navigation ies. Julia has moods and caught and takes players to certain attitudes, overnd, she has a good idea. She recalls what players told her what they did to do when they met this last, etc. (Transcripts of interactions with Julia . It is an example of what Maudlin is Chatterbot named. A chatterbot has different modules to deal with the different functions necessary to automate a player in a mouth. The conversation module is implemented as a predetermined layer of mini experts consisting of a set of behavior patterns and the associated possible responses. Chatterbots go over the program Eliza Beyond wheat tree, as they have a coarse number of "Trick" use and over a highly developed reminder of past events, conversations, etc. feature.
To design these entertainment agents, the same basic questions must be answered, which are centrally located for any development in the field of agents, namely perceptual, selection of behaviors, motor control, condition and communication. The agent must perceive its often dynamic and non-predictable environment, especially if it is possible for a user to charge them. The Agent MUB decide what to do as a nesting so that there is progress for those tasks, to whose solution he has been designed. The corresponding acts must be converted into concrete motor commands. Over time, the agent’s behavior on the basis of its previous experience others and improve. Finally, the agent will be able to communicate with other human or artificial agents in the world.
The key problem is to find an architecture that includes all these functionalities and leads to a behavior that is fast, reactive, adaptive, robust, self-contained and last but not least lifelike is. Lifelike behavior is non-mechanical, non-predictable and spontaneous. The architectures of many of the caught successful entertainment agents show an amazing number of similarities ::
- The agents are built as distributed, decentralized systems consisting of simple competence modules. Each competence module is one expert for the execution of a certain, simple, task-oriented activity. There is neither a central thinking nor a central internal model. The modules are in contact with each other through extremely simple embassies. Each of the competence modules is connected to corresponding sensors and effectors. As a result, the generated behavior is robust, it adapts to changes and is fast and reactive. Complex behavior is the result of dynamic interactions (back-coupling loops) at three different levels: the interaction between agents and the environment, between the different modules within the agent and between the different agents. A simple Braitenberg manhole, for example, the user like, can be constructed so that it moves to him at a speed that is proportional to the distance from it. As an example of a complex interaction of many agents, Reynold’s dramatically demonstrate the herd shoot through simple local rules that are followed by any animal of the herd.
- The architecture includes a variety of redundant methods for the same competence. Many complexity and intelligence layers provide for fault tolerance, elegant degradation and behaviors that do not work mechanically. For example, Julia has different methods of the answer to exercises that have been addressed to them: it can try to understand the exercise and generate a correct answer, or it can, if that is missing, cite a different way to the same topic or if that is missing, starting with another topic.
Apart from the standard research questions, the construction of entertainment agents requires the procurement with novel questions (especially with those of artificial intelligence and certainly with those of artificial life), z.B. How to get hatching, intentions, social behavior and entertainment. Typically, these questions are more important than the task of making the agent particularly intelligent, there to quote bates, "The actual challenge is to create a consistent apparent of beweed, intention and social relationship."
Although these topics may be of crucial importance in the construction and understanding of intelligence, they have hardly been studied in the artificial intelligence. The construction of agents for the entertainment area forces the researcher ultimately to deal more with the user. The researcher is forced to work with the user’s psychology: How is the typical user perceiving the virtual figures? Which behavior will show? Which mIB messages and confusing situations will occur?
Other subjects such as human-computer relationships, animation, sociology, literature and theater are particularly helpful in answering these questions. Animation teaches us, for example, that the typical user figures moving faster than young, better and intelligent drawstring. Literature and Theater teach us that it is easier for us to capture stereotypic figures than the Shrink in program Eliza.
The Alive project
The detailed description of a particular project, which is aimed to build entertainment agents, may possibly explain the challenges for research and the application possibilities of the entertainment agents in a convincing manner. Alive is a virtual environment that allows wireless whole body interaction between a human participant and a virtual world inhabited by animated autonomous agents. Alive is called "Artificial Life Interactive Video Environment". One of the goals of the ALIVE project is to show that virtual environments "emotional" and convey more impressive experiences because the participant can interact with the animated figures. The ALIVE system was presented and tested at various public events.
In the style of Myron Krueger’s Videoplace The ALIVE system offers unrestricted full corporate interaction with the virtual world. The user of Alive moves in a room of about 16 x 16 Feet. A video camera starts the image of the user and builds it into a 3D world after being out of the background. This image is projected on a coarse screen, which is the user opposite and as a kind magical mirror Works: The user sees itself, surrounded by different objects and agents.
For the interaction with the virtual world, neither glasses, gloves nor cables are necessary. Computer vision techniques are used to obtain information about the person, z.B. over its 3D position, over the position of the different body parts and via various simple movements. Alive combines active seeing and expert knowledge to realize robust real-time presentation. The location of the user and its arm and body movements influence the behavior of the agents. The user receives visible and horny feedback from the internal creation and reactions of the agent. The agents have internal requires and motivations, various sensors to perceive their environment, a repertoire of actions that they can export, and a body-related motor system that allows them to move in the environment and act in it. A behavior program decides in real time which acts of the agent exports to satisfy its internal needs and to respond to the possibilities of the current state of its environment.
The Alive System. The dog runs in the direction in which the user shows.
The system not only allows the obvious, directly influenceable type of interaction, but also a more powerful, indirect nature of the interaction in which the movements can have more complex meaning. The importance of a movement is interpreted by the agent due to the situation in which he and the user are located. If the user shows, for example, in a certain direction and sends the figure thereby, the figure moves to another place in the environment, depending on which the user is (and in which direction it shows). So a relatively small number of gestures can mean many different things in many different situations.
The ALIVE system contains a tool called Hamsterdam to build half-intelligent autonomous agents that interact with each other and with the user. Hamsterdam generates agents that respond with a relevant action at every time step, as a function of their internal needs and motivations, their past, the perceived environment and their inherent opportunities, challenges and changes, with a relevant action. In addition, the patterns and rhythm of the chosen actions are so that the agents do not change between multiple activities, remain in a single simple activity.
You can interrupt any activity when an urgent requirement occurs or results in unforeseen possibility. The Hamsterdam action model is based on behavioral models of animals proposed by conductors. In particular, various concepts suggested by conductors, such as behavioral hierarchies, exclusive, encouragement, etc. shown that they are crucial for robust and flexible behavior that is necessary for the autonomic interactive agents. The Alive system shows that animated figures that are based on models of artificial life can convincingly act (D.H. You switch out doubt).
If Hamsterdam is used in the construction of an agent, the designer determines the sensors of the agent, its motivation, its inner needs, its activities and actions. With this information, Hamsterdam automatically decides which activities are most important at what time for the agent, according to its condition, the situation in which he is located, and his behavior immediately before. The observed behavior or the acts of the agent are the result of numerous actions that compete for the control of the agent. They compete on the basis of the value of a given activity of the agent at a certain moment, according to the perceived environment, the internal needs of the agent and the immediate past of the agent.
The ALIVE system consists of various virtual worlds between which the user can change by printing a virtual knob. Each world is inhabited by different agents: A world is inhabited by a doll, a second of a hamster and a bird of prey and a third of a dog. The doll follows the user (in 3D) and tries to hold his hand. She also imitates some user activities (sit down, jump, etc.To). She removes when he moves away and comes back when he waves. The doll uses various expression to show your inner condition. For example, she polls when the user sends them away, she laughs when she is retrieved, and she cichles when he caught her belly. The hamster bypasses the counterpart, follows the user and begs food.
The hamster lays down on the jerking to let her abdomen crawl when the user bent down to stroke him. If the user caressed the hamster for a while, his needs for attention is filled and another activity is given priority (Z.B. Dietary. The user can feed the hamster when eating food from a virtual table and laying it on the floor. The user can enter the bird of birds out of the Catige into the world of the hamster. The bird of prey tries to hunt the hamster and dead him. The bird bird stops the user for a bird of prey and tries to avoid him and flee in front of him. The hungry he becomes, so the coukhner he becomes, and dares to get to the user closer to come. Both the bird of birds and the hamster can tune their different inner computers (avoid the bird of prey, find food, not in obstacles, etc.To).
The most intelligent figure that has been developed so far is the dog Silas. The repertoire of Silas extends to the following behaviors so far: it follows the user, sits down (if the user agrees), goes away (if the user sends it away) and leads out other tricks. He jumps, picks up a ball, lay down and shake up. He also hunts the hamster when he is introduced into his world. In addition to the visual sensors and feedback, simple toneins and outputs are used in the dogworld. The user opposes a direction microphone. The resulting signal is supplied to a simple pitch tracker. High Tone (Z.B. Handlocks, whistles, shrill voices) and deep tones (soft voice) are interpreted as positive and negative inputs from the user to the dog. The dog also reflects horny outputs that consist of a series of entered patterns.
- 1. By observing thousands of users interacted with the agents of Alive, a lot has been learned. The gestures that the user can export should fit intuitively in the dians and provide immediate feedback. For designers of user interfaces the latter will be self-resistant, for the researchers in the field of artificial life and the artificial intelligence it is not. Examples for natural Gestures are caresses of animals or send forward and waving for the virtual doll. Whenever an agent has successfully recognized a gesture, the user should immediately receive feedback either in the form of movement and / or as a facial or body expression (the hamster rolls, for example, on the jerking when it is caressed, the doll Laughters when you are tickled, etc.To). This helps the user to develop understanding for the scope of recognized gestures.
- 2. Even if the gestures of the environment are adapted, it is necessary to have a human Leader to have, which gives the user hints what he could do ( "Try to stroke the hamster", "The doll will go away when you send it away" etc.To). The current ALIVE system has an artificial driver, which has been introduced as another autonomous agent to fulfill a special role: he observes the user, stores which interactions the users with the world has and is occasionally proposing, he speaks or serves the appropriate gesture. The fuhrer in the virtual world is a parrot because one expects a parrot that he speaks, but not that he understands language.
- 3. Users are tolerant against imperfect behavior of the agents (in contrast to the counter) such as delays and occasional wrong or incorrect recognition. The presence of agents prompted people to put appropriate expectations of the performance of the sensor system. We have learned that people expect that virtual, immovable objects reliable work, D.H. The reaction of the object MUB immediate, predictable and consistent. On the other hand, people accept that animals or human agent have perceptual dermogments and are in a state, and can accept that the agent does not notice something. Therefore, gestures that are difficult to recognize, such as. B. Waving, successfully used in connection with agents (an agent may have not seen that the user beckons). But the same gesture was frustrated by the user when used with immovable objects, Z.B. a switch.
- 4. It is important to make the motivation and the creation of the agent in the external facial events of the agent visible. A sophisticated figure such as silas the dog becomes with his eyes to lead, D.H. He turns to look at an item or a person before he goes to the subject or humans (for example, to take the subject or to call the person to play with him). If a figure does not lead with her eyes, so her behavior works mechanically and therefore not very lifelike. Motivation and sawing must also be made visible because the user if he can not recognize the inner variables, which determine the behavior of the agent, will possibly be probably confused or obvious. For example, if Silas is hungry, he may not obey. So it is important that the user recognizes that Silas is hungry so he understands why Silas reacts differently than a few minutes ago.
- 5. The crucial lesson consistently insisted that it is not so important for the design of an attractive immersive environment, as impressive the graphics, but how meaningful the interactions that the user can export. Alive users have enhanced that they had a coarse pleasure with the system and the interactions with the figures. Especially they seem to skate worlds that of fouling Agents are inhabited, with which the user can take a feeling relationship. For example, the users of the doll were very pleased. She felt bad if her actions prompted the doll to paste, and she was happy when the doll laughed.
The Alive system proves that the area of entertainment can be a challenging and interesting area for AA research. Alive supplies a new environment to study architectures for intelligent autonomous agents. As a prud for agent architectures, it avoids the problems arising with real hardware agents or robots, but at the same time, it forces us to commit our problems such as loud sensors and an unpredictable, fast-changing environment. ALIVE allows us to study agents with a higher level of knowledge enhanced, without having to simplify the world in which the agents live are simplified.
Alive is only the beginning of a series of novel applications that can be explored with such a system. At present, we examine Alive to the possibility of application in the area of interactive storytelling. The user plays a figure in history, and all other figures are artificial agents who work in that history develops. Another obvious field of application of Alive in the field of entertainment are video games. We have connected the ALIVE’s view-based interface with existing video games so that the user can control the games with his entire body. In addition, we examine how autonomous video game characters can learn and over time improve their skills so that the game remains interesting for the player. Finally, we develop animated figures that teach the user a corpulent ability in a personalized way. The agent is designed as a personal trainer who shows the user as he has a activity, and he gives the user a personalized and temporal feedback, based on the sensual information about the gestures of the user and its body positions.
Recently, systems such as the vigor, Neurobaby, the Fish of Terzopoulos, Julia and Alive show that the area of entertainment is a pleasurable and challenging field of application for AA research. At the same time, these early experiments also show that this field of application requires an interdisciplinary approach that combines insights from the humanities with the computer models developed in artificial intelligence and artificial life.
The contribution of Patti Maes "Artificial Life meets Entertainment: Lifelike Autonomous Agents" is first in the Special IE on New Horizons of Commercial and Industrial AI, Vol. 38, NO.11, Communications of the ACM, November 1995, published.
Literature from the English translated by Florian Rotzer