A multi-agent system for information architecture

A MULTI-AGENT SYSTEM FOR INFORMATION ARCHITECTURE By Jorge Alberto Lopes Gil Placement student sponsored by British Telecom. Bartlett School of Graduate Studies University College London September 2000 KEYWORDS: Evolutionary Architecture, Systems behaviour, Artificial-life, Software Agents, Information Visualization A BS TR A C T Most communication between people nowadays goes through the channels of electronic information networks. How can this network space be designed to facilitate access to information? Can architecture as an art of designing spatial interfaces play a role? How will architectural design be affected when moving into the information medium? A scenario of a networked office environment is used to explore the problem. I propose the use of a three-dimensional virtual environment to register information flow and usage. This scenario points to a new kind of architecture based on information visualization and with an evolutionary character. I design a multi-agent system to manage the virtual environment. The emergent behaviours that result from the system are described and further explorations pointed out. The result is an information landscape that can be explored by the user to access data that flows on the communications network. 2 TA BL E OF C O N TE N TS 1. Introduction ..................................................................................................................... 4 2. Multi-agent systems: two examples ................................................................................. 13 3. A multi-agent system for information architecture ............................................................. 15 3.1 Programming environment.................................................................................... 16 3.2 System architecture .............................................................................................. 17 3.3 System dynamics ................................................................................................. 20 3.4 Emergent behaviours ............................................................................................ 21 4. Possible explorations ....................................................................................................... 23 5. Conclusions .................................................................................................................... 30 6. References...................................................................................................................... 32 Appendix – Log of the project development .......................................................................... 58 3 “Our electronically-configured world has forced us to move from the habit of data classification to the mode of pattern recognition. We can no longer build serially, block-by-block, step-by-step, because instant- communication insures that all factors of the environment and of experience coexist in a state of active interplay.” Marshall McLuhan, The medium is the massage fig. 1Screen of code from ‘The Matrix’ 1 . IN TR O D U C T I ON The Internet is a network connecting more than 300 million users. This population communicates through the network producing a constant flow of information [fig.2]. They use it for social, educational, research, commercial, work, entertainment, etc. activities. As these activities on the Internet intensify, more and more information becomes available and is exchanged, increasing the relevance of this medium in all aspects of human life. Modern communication channels increasingly render the conventional transmission of documents, letters and other information obsolete. This way most communication between people goes through the electronic information networks. fig. 3 Internet traffic visualisation – S. Eick fig. 2 Worldwide distribution of a company fig. 4 Videoconference on desktop computer If we consider the effect on global economy, we see how companies gain the ability to leave their exclusively local influence/presence. Multinationals enter every market and open branches worldwide [fig.3]. The new electronic media allow remote offices of the same company to maintain a closer relation improving the company’s identity. Meetings take place without the physical displacement of people, in the virtual space of the networks [fig.4]. Occasional collaborations between companies can also be established for different reasons. For example to take advantage of the different time zones in different areas of the globe, allowing a project to be developed around the clock, or to share specialized human or technological resources exclusive to a certain company. [Mitchell, 1995, pp. 92-98] A sc e na ri o In a networked work environment documents about ongoing projects are exchanged and stored in servers, accessible from every node of the company’s network. Traditionally this information is kept under a tree hierarchy of folders organized by the users or an administrator and explored using text based tools [fig.5]. At a certain point, when the amount of documents becomes very large, it is difficult to navigate in this hierarchical structure, cross-references are difficult to handle and a single file may get lost if placed on the “wrong” branch [fig.6]. Visualization 4 enables the communication of large amounts of information to the human visual system [Girardin, 1996]. Moving to threedimensional representations is a natural step. In a work environment a three-dimensional information space has the role to register information flow and usage for a particular project. When a certain document relating to the project is created on the network, an event is sent to the information space that processes it creating a register for the user to consult. Such a registration is a living process that absorbs a constant information flow [fig.7]. The virtual environment handles with the address of the files on the network and with a set of identifiers that allow differentiation and interpretation of the data. References can change location, be duplicated or simply removed without interfering directly with the file structure on the network. The user, by navigating this virtual environment, can have access to documents related to the project in a spatial experience [fig.8]. This scenario suggests a new kind of information visualization based on three-dimensional representations with an evolutionary character. How can this network space be designed to facilitate access to information? Can architecture as an art of designing spatial interfaces play a role? How will architectural design be affected when moving into the information medium? These are some relevant questions addressed by this project. fig. 6 BSCW server screen of project fig. 5 Usenet crosspost visualisation – Netscan by M. Smith fig. 8 PITS: Populated Information Terrains – D. Snowden, U. Nottingham fig. 7 3D model of vBNS network – J. Brow at Moat Ob j e c ti v e s The objective with this project is to create an object of evolutionary architecture that exists inserted in the information territory of Cyberspace, like the office network environment (Intranet), to become an interface between users and the flowing streams of data. This architecture will be designed as a multi-agent system with every element (information, architecture object, user) having an individual character and role. The architectural object is embedded with mechanisms that link it to the existing flows of information. An adaptive system is a “system with the capacity to modify its internal state or structure in response to changes in environmental demands or opportunities.” [OSG, 1981, p.17] 5 This is precisely the main characteristic of this evolutionary architecture: a virtual environment responsive to the permanent changes and events generated in the information territory. The result will be what I call architectural life, a particular form of artificial life, which focuses on the creation of living artificial environments. The initial structures will change over time as the agents interact with each other. The information will self-organize in an architectural form. The number of interacting agents and the complexity of these interactions can give rise to a web of structures - social, spatial, contextual and architectural - generating some kind of community ecology, considered as the assemblage of interacting agents of different types operating in a specific environment. The chain of actions and reactions between agents drives the system into a cycle similar to the cycles of an ecosystem. In the next chapters I approach first the subject of information architecture, where ideas from evolutionary architecture find its natural ground through the relation to cybernetics and Cyberspace. Then the subject of living systems focusing on artificial ecosystems and multi-agent systems, followed by a description of the developed system and some preliminary results. fig. 9 Movement of a 3D Cellular Automata Surface, using CAPOW by Rudy Rucker Architecture is the interface. Information is the raw material. Cyberspace is the environment where all the events take place. 2 . M UL T I A GE N T S Y S TE MS : T W O E XA M PL ES fig. 10 Mobile robots playing football After understanding the principles behind the agent - the individual component of an artificial ecosystem - the next step is to understand the principles behind the whole living system – a multi-agent system. 6 A multi-agent system is a complex system that generates special dynamics using many agents with simple behaviours [fig.36]. The collective interactions allow complex behaviours to emerge. It can have many agents of the same kind and/or different kinds of agents. 2.1 ECHO Echo is a model of an agent based ecosystem, developed by John Holland, “father” of genetic algorithms. He developed this model to present a theory of complex adaptive systems (CAS). Like other complex systems it presents emergent behaviour resulting from the non-linear interactions between agents. But it is called adaptive because the agents adapt their behaviour to other agents, causing the system to evolve over time. Adaptation is in biological usage the process whereby an organism fits itself to the environment [Holland, 1995, p.9]. Adaptation in Echo happens through rule discovery, in a process that builds and evolves the rules from a simple set of elements, that he calls schemata or building blocks. A genetic algorithm procedure produces the new rules and provides for its evolution through selection. For my project the relevant aspect of this model is the agent architecture [fig.38] and the methods of interaction: the fact that it is a complex system. The agents use an input/output device to interact: detectors to receive stimuli and effectors to act upon the environment or other agents. The capabilities of individual agents are what let us understand the interactions of large numbers of agents. These capabilities can be determined by a collection of simple stimulus-response rules for processing the received information. These rules are defined by an if/then/else structure. The agents’ capabilities in a fixed point in time are fig. 11 Architecture of agents in ECHO fig. 12 The world of ECHO defined by these three elements: detectors, effectors and rule set [Holland, 1995, p.88]. The major part of the modelling effort goes into selecting and representing stimuli and responses, because different selections result in different models. The selection varies according to the questions that are being investigated. [Holland, 1995, p.7] The agents in Echo have another particular feature. They have an identifier called tag. Tags are a pervasive feature of CAS because they facilitate selective interaction, allowing agents to select among agents or objects in the environment [Holland, 1995, p.14]. 2 . 2 S tar l o g o 7 Starlogo is a programming language that enables the easy design of multi-agent dynamic systems, taking benefit of its parallelism. The examples created with this tool allow a clear understanding of how decentralized, self-organizing communities evolve. When building new systems people tend to use centralized strategies. When there’s a need to have control over the result such strategies are very useful. Other times it happens due to the implanted centralized mindset in our society, based on rigid hierarchies, where a leader defines the development of the system [Resnick, 1998, p.4]. This is the so-called top-down approach. By using multi-agent systems we obtain an emergent behaviour not programmed in the strategies of the individual agents [fig.39]. The global strategy emerges from the local interactions of the agents, which have no conscience of the global state of the system. This is the bottom-up approach. fig. 13 Termites piling up woodchips in Starlogo A relevant feature that distinguishes Starlogo from other artificial life ecosystems is the importance given to the environment. The environment plays a role in the overall behaviour of the system, because it’s seen as something to interact with, thus influencing and constraining the behaviour of agents. [Resnick, 1998, p.143] In Starlogo it is equivalent to other agents having a set of rules that determine its local behaviour. Communication is possible between agents and environment, increasing the levels of interaction in the system. It also allows, “distributed cognition” [Resnick, 1998, p.34] by delegating tasks of mobile agents to the environment [fig.40]. This is an important feature when my focus is on considering the elements of the virtual environment - the architecture objects - as agents that have behaviour and participate actively in the dynamics of the system. fig. 14 Slime-mold cells aggregating using pheromone passed to the environment 8 3 . T H E M UL TI -A G EN T S Y S T EM F OR I NF OR MA TI O N A R CH I T E C T UR E The project is based on a multi-agent system (MAS) using software agents to manage the flow of data and build a virtual environment. It has information agents that deal with the management of data, selecting and distributing it in the virtual environment, built by architecture agents that receive data from the info-agents responding accordingly. The virtual environment is an information landscape to be experienced by the user, giving him access to the data content. This information landscape is an architectural object with a living character, whose dynamics rise from the interactions of the multi-agent system. [anim.1] The developed system is only a simulation of a real scenario. Several random values feed the system’s parameters with the required data to make it work. When inserted in a real network environment, real-time events will start and feed the system. 5.1 PROGRAMMING ENVIRONMENT The MAS is programmed and designed in VRML (Virtual Reality Modelling Language) and JavaScript, using a VRML code editor (VRML Pad 1.0) in a Microsoft Windows environment. fig. 16 VRML browser interface (Cosmo Player 2.1), integrated in Internet Explorer 5.0 5 . 1 . 1 Pr o gr amm in g l an g ua g e s VRML is a descriptive language for designing three-dimensional environments navigable in real-time over a network. Using an Internet browser with a VRML plug-in the user interface and the navigation are immediately available [fig.42]. As designer it 9 gives an advantage over other three-dimensional computer graphics systems, because I can be concerned solely with building the geometry of the world and behaviours of the system. JavaScript is the easier scripting language to implement behaviours in VRML worlds. The behaviours play a fundamental role in this project since it is a dynamic system with individual agents with individual behaviours and about the interactions between these agents. The use of JavaScript is intensive in the program and uses most of the time spent in the development. An annotated version of the source code can be found in Appendix B. 5 . 1 . 2 S y s t em p ri nc ip l e s The selected programming languages allow the implementation of two fundamental principles of a MAS: individual based modelling and agent communication. VRML has a PROTO node that works like a class in Object Oriented Programming (OOP). This means that an object is created with an internal definition and several independent instances of it can be inserted in the virtual environment. A MAS is an individual based model in which each agent is an autonomous individual that has a behaviour defined by a simple if/then rule set. I use the PROTO node to create an object for the information agent and another for the architecture agent. The character of each agent is defined in this object by a behaviour script with the rules encoded. Then I can insert as many of these agents as necessary by invoking the corresponding PROTO. This is the way to obtain multiple agents (individuals) acting in parallel. Furthermore each agent is autonomous meaning that although they share the same behaviour rules, they can do different things at the same point in time, depending on the local circumstances. This becomes possible through agent communication [fig.43]. fig. 17 Diagram of PROTO use in multi-agent systems VRML allows the exchange of data between objects by creating Routes that send an Event Out from one object to the Event In of another. Agents use these connections to send data or simple stimuli, triggering actions on other agents. 5.2 SYSTEM ARCHITECTURE 10 fig. 21 User Interface connections with Black box fig. 18 Diagram of the System Architecture with the different components and their connections. Now I move to a description of the system’s architecture in terms of diagrams that go from the large scale of the global system [fig.44] to the more detailed scale of the behaviour in individual agents. 5 . 2 . 1 D a tas pa c e The Dataspace is the information context where the MAS is integrated. In the proposed scenario it’s a company network, where documents and messages are exchanged between users or user groups. The Dataspace sends events into the MAS whenever some activity occurs related to the specific group that the system represents. In this model a library of pre-selected documents that are randomly picked simulates this procedure. The Dataspace is restricted to a folder with files that can be referenced by their “url” (internet address) and is used as an example of a real Dataspace. For demonstration purposes the file formats have been divided in text documents, images, sounds and urgent messages. fig. 20 Information agent’s connections with Black box fig. 19 Architecture agent’s connections with Black box 5 . 2 . 2 Bl a ck bo x Black box is defined as “a (component of a) system that is only considered in terms of its inputs and outputs. Its internal mechanisms are unknown or ignored.” [OSG, 19881, p.17] The Black box is the central unit of the virtual environment, where the geometries of the scene graph are structured and all the pieces that compose it brought together. The Black box doesn’t have geometries or behaviours, only some general environment settings related to navigation, viewpoints, background colour, lights and sensors to locate the user. 11 The second function of the Black box is to connect all the agents together, working as a communications central. It redirects the incoming events from the agents that arrive at its inputs to the relevant outputs that feed forward the events. Each agent has a bi-directional link to the Black box sending and receiving stimuli [figs.45-47]. 5 . 2 . 3 U s er In t er f ac e The user interface establishes a connection between the user and the MAS by linking him to the Black box. The interface doesn’t have any sort of behaviour like the agents, because it’s linked to the only intelligent agent of the system with an implicit behaviour: the user. Its main function is to provide the user with information about the status of the environment, displaying it augmented reality style over the scene [fig.48, anim.11]. The user has direct access to the data contained in all software agents by selecting them with the mouse. He learns about the name of documents when passing with the mouse over an agent, and sees icons that represent documents or gains access to “urls” that open the real document when clicking with the mouse on the agent. fig. 22 Picture of the Info-source PROTO fig. 23 Example of data being displayed on user request 5 . 2 . 4 I nf or m at i on so urc e The information source (info-source) is the software agent responsible for receiving the information from the network and passing it to the information agents (info-agents). In this model a grid and a cube represent it [fig.49]. The grid is an area that defines the limit of movement of the info-agents, and is the same size as the information landscape. The cube is located in the centre of the grid and represents the actual source of information where the info-agents collect data. The data is encoded in a tag composed by: a colour representing the type of document, a name with a description of the document, an “url” with the location of the document on a server and an iconographic representation of this document. 5 . 2 . 5 I nf or m at i on ag e nt s 12 fig. 24 Info-agents at “work”: 1- seek state, 2- empty state, 3- selecting area The information agents (info-agents) are responsible for bringing the information content into the virtual environment. They are represented as coloured spheres that move within the info-source’s grid and have two states: “seek” (coloured, when carrying data) or “empty” (white, on the way to collect new data) [fig.50]. The colours allow visual identification of content flow and a global analysis of agent activity. Some patterns and behaviours can only be noticed with the apparent difference between agents. The difference is only apparent because they all have the same behaviour rules. When the system starts running all the info-agents initiate their activity from the source location. They are set in “seek” state fig. 25 Detailed diagram of info-agent, with internal structure and operative connections and assigned a tag. The tag is what the info-agent uses to communicate with other agents, including the user. The user can at any time use the mouse to have access to the info-agent’s load. After leaving the source the agent selects a target on the grid and moves in that direction. When reaching the target it enquires the architecture agent for the local type of information. If the location is empty, or the content is of the same type as the infoagent’s load, a transfer action is triggered. It transfers the tag to the architecture agent sending the corresponding events. But the location might have information of a different type. In that case the info-agent selects a new target on the environment, and proceeds to drop its load [fig.50]. The cycle repeats until it is able to reach a location that satisfies one of the two previous conditions: to be empty or with information of the same type [anim.3]. These two conditions make sure that the agents organize the data with a minimum of criteria. 13 After transferring the tag it automatically changes to empty state and returns to the source. Reaching it the agent collects the tag of a new package of data. 5 . 2 . 6 A rch i t ec tu re a ge n t s The elements that build the information landscape are agents like the info-agents with a set of rules for local action. They are fig. 26 Neighbourhood association by arch-agents. The active ones display a link and mark the empty neighbours with the same data id. The grid boundaries don’t have archagents responsible for storing data and offer the user an environment that represents the flow of information through the system. They are contained in a white plane that constitutes the initial surface of the information landscape [fig.52]. An invisible grid structures the plane with each node having independent behaviour. This grid is parallel to and has the same size as the infosource’s grid. The parallelism is helpful when comparing the activities of both worlds: information and architecture. When the system starts all agents are set in their initial state: elevation is “0” and the tag is empty. This only changes when the info-agents send a stimulus with a tag. On receiving this tag the arch-agent adds a value to the elevation, assigns the tag to itself and communicates with the eight surrounding neighbours [fig.53]– the Moore neighbourhood [Toffoli, Margolus, 1987, p.60]. It assigns the same type of content to any neighbour that had no type assigned previously. The arch-agent that initially received the tag displays a link. The link is a coloured text string that identifies the type of content of the arch-agent and it always reorients itself to face the user. These links can be activated by the user, which receives via user interface a list of the “urls” and icons of the different documents contained in the arch-agent. fig. 27 Asynchronous sequence of events. They start from the info-source and cycle through agents eventually reaching the user. fig. 28 Detailed diagram of arch-agent, with internal structure and operative connections. As info-agents move around distributing data the elevation of the arch-agents keeps changing turning the initial plane into the information landscape with links giving access to its content. But this description is only of a one-way adaptation. The more data the architectural surface receives, the more it changes [fig.54]. 14 However the arch-agents have a reverse behaviour that increases the dynamics of the MAS towards a living system. Each arch-agent has a timer that starts countdown when it receives data. The only thing that resets the timer is the user activating the link of the arch-agent. After a certain period of inactivity, whenever the timer passes its cycle value, the arch-agent reduces the height value. And if the height value reaches “0” it restores the initial state: the tag is emptied. 5.3 SYSTEM DYNAMICS The MAS is dynamic because of the interactions established between the different agents. In a parallel system involving multiple simultaneous actions it becomes difficult to grasp the overall dynamics. There’s a sequence of steps when data enters the system coming from the Dataspace that represent these dynamics [fig.55]. After reaching the info-agent it starts a dynamic cycle. The user is responsible for retrieving it from the system by direct action on the agents. There are two types of flows involved in these dynamics that keep the systems integrity. One is the information that flows fig. 29 Initial state of the information landscape, and after some info-agent activity through the system and the other is the time flow. 5 . 3 . 1 I nf or m at i on f lo w The system presents two information flows: an external and an internal [fig.56]. The external flow consists of the information that enters through the primary input (info-source) and after being processed fig. 30 The two information cycles of the system leaves the system on user retrieval. The internal flow is that of information being processed within the system. The exchange of information between agents is the indicator of the system status at any given moment, and eventually of its living status. The internal cycle is responsible for the systems unity as an organism. It represents a dependence chain of all the constituent agents, and the removal of one would break the chain of interactions and cause the others to cease their activity. Each agent plays a unique role in the system. 15 5 . 3 . 2 T im e fl ow In a parallel system time doesn’t follow a single linear arrow, it is fragmented into the cycles of each individual. Time appears in two formats: continuous and discreet [fig.57]. Apart from the global system that follows the arrow of time in its continuous adaptation, the individuals create new arrows with their dynamics. The continuous time is internal to each single agent, like the heartbeat of an organism, controlling the individual behaviour. fig. 31 The two time formats of the system: Continuous time inside the agents, discreet time in the black box. Continuous time flows in a regular way and is responsible for the movement of info-agents and the reverse behaviour of the arch-agents. fig. 32 The different viewpoints: two overviews of the surface, navigating the surface, front view of the two worlds, two overviews of the info-source, navigating the info-source, and between worlds. The discreet time is irregular and is external to the agents. It’s the time of the Black box and is defined by the rhythm at which events occur. 5.4. EMERGENT BEHAVIOURS The emergent behaviours produced by the described system are still very elementary in terms of living systems. Most are related to mobility aspects of the info-agents and not to interactions between agents: they don’t produce a corresponding pattern in the information landscape because this interaction is related to information placement and not agent movement. But some are typical of multi-agent systems and future improvements can increase the complexity of these behaviours. 16 There are several viewpoints in the system to help the user navigate to particular positions giving better observation of the behaviours. These are “overview”/”top view” of the information landscape and info-source, or he can be set right on the information landscape or info-source to navigate among agents [fig.58, anim.2]. fig. 33 The circle pattern around the info-source 5 . 4 . 1 T he cir cl e p a t t er n The circle pattern is a movement pattern of the info-agents and the first to emerge. When the system is started all the infoagents move away from the centre forming a circle that grows until some start leaving the “formation” after dropping the data they were carrying and return to the info-source [fig.59, anim.4]. This is a typical behaviour that comes from having many agents starting from a single location with random destinations [Resnick, 1994, p.96]. In this case the circle “blooms” showing initially a irregular shape, because all agents start with the same direction and update it progressively towards their target. 5 . 4 . 2 T he fu gu e fig. 34 The agents moving away from the grid… and about to return The fugue behaviour results from a curious Windows environment feature. If the browser window is minimized for some time, when restoring its size the agents are gone…but after some time you see how they are back in the info-source grid [fig.60, anim.5]. The info-agents movement is direction driven and not position driven: it is continuous and the direction updates towards the target. If the window is minimized the current direction is kept and for some reason the movement doesn’t stop. When the window size is restored they restart updating their direction towards the previous target. So what we observe is a sudden return of the agents to the info-source grid, as if coming back to work after being left unattended. They immediately restore their activity. 5 . 4 . 3 T he co rn er m e e tin g & w a ve p a t te rn s 17 fig. 35 The agents gathered in the corner of the grid The previous behaviour was fixed to a certain extent by constraining the info-agent movement to the grid area. They are instructed not to move beyond that grid limits. But like before they don’t stop when the window is minimized and they follow the grid limits trying to move away from it. After a certain time, when the window size is restored they are gathered at the corners of the grid and restart updating their direction [fig.61, anim.6]. As a result a new pattern emerges: a quarter of a circle grows from each corner. And as the quarters meet in the centre a circle grows designed by all the info-agents that were in “empty” state. The overall aspect is that of waves in a container and overlapping. 5 . 4 . 4 T he “ air tr a ff ic ” After running the program for some time it can be observed how info-agents get locked in a continuous circular movement around a point of the grid or around the source [anim.7]. And suddenly they leave this behaviour and return to the normal activity. I called it the “air traffic” behaviour. fig. 36 Diagram of the event stack procedure, with the waiting agents progressively leaving on Event out. This behaviour occurs because the MAS is parallel with individual agents sending events to a single processor. An event stack collects these events as they arrive, stores them with a timestamp and they are processed in that order [fig.62]. As a result the info-agents have to wait for their turn, starting to circle around the target or the source until they get “permission” from the Black box to proceed, in the form of the expected event. If they were waiting to drop data on an arch-agent they turn “empty” and move to the info-source. If they were waiting to collect data from the info-source they change colour and move to the new target, away from the centre. 18 fig. 37 Sequence of landscape modulations 5 . 4 . 5 T he su rf a c e b re a t h The surface breath is a direct effect of introducing a reverse behaviour on the arch-agents. They have a continuous heartbeat that allows them to change their parameters independently from the discreet stimulation by info-agents. As a result the modulation of the information landscape that they build isn’t constant: it breathes like an organism [fig.63, anim.8]. The hills on the landscape grow and shrink as a reflex of the dynamics of both info-agent and arch-agent activity. This effect becomes more apparent if the heartbeat speed is high, increasing arch-agent metabolism. 5.5 POSSIBLE EXPLORATIONS fig. 38 The two worlds where interaction is established: architecture and information The big reward comes from observing the system and watching things come alive. It is the best way to develop intuitions on how decentralized MAS’s work [Resnick, 1994, p.5], thus improving the created models. And the resulting surprises can show ways of research previously not considered. Several suggestions follow, for experiments that can be made with this particular system that resulted from the active observation of it running. They are intuitions or speculations into what can be tried out to move the system closer to the objective of having a living organism. They introduce new levels of interaction that improve the dynamic relations and the complexity of behaviours between the two worlds [fig.64]. 5 . 5 . 1 T he us e r a s a t tr a ct or The user can become an attractor in this MAS being the first target chosen by info-agents. The user has the freedom to move around the landscape while consulting the data links. This interaction is recorded by the arch-agents preserving the areas more frequently visited and eliminating the hills of abandoned or forgotten areas. The info-agents will always move first to the user location and add new data if it is of the same type. This way the information that is currently interesting to the user will 19 concentrate faster in his location. And the other types will be placed and replaced in his vicinity, as incompatible info-agents move away looking for a free area. This improvement increases the power of the user indirectly (implicitly) influencing also the global system’s evolution. 5 . 5 . 2 D if f er en ti a t ed In f o -a g en t s The rules define the character of all agents producing the individual actions that define their role in the system. Playing with the parameters of the rules changes the character of the agents creating new types of individual behaviour. These new behaviours will consequently determine new developments of the system and the emergence of new collective behaviours. Some editable parameters are: the number of agents in the system, the speed of movement of info-agents, location and quantity of infosource, rate of change and amount of change of arch-agents. A further step is to introduce differences between agents of the same type, adding diversity to the system both at organism level (info-agents) and environment level (arch-agents). New behaviours and patterns will emerge from the richer community ecology. 5 . 5 . 3 R a nd om m ov e me n t o f in f o -a g en t s The info-agents can be assigned a free random movement when in “seek” state. Instead of being driven towards a randomly picked goal, they scan the arch-agents for data with the same type as the one they carry. This requires more frequent communication between the info-agents and the arch-agents, inquiring about the local state of the surface, thus slowing down the system with the higher traffic of events. But on the other hand the agents start behaving in a more natural and local way using something similar to vision. They stop guessing a remote location and acting on a trial and error basis. By introducing this character in the mobile agents, the way information is stored changes significantly. It shall present a higher concentration of data around the info-source, since they always depart from there in “seek” state, and as soon they come across the first cluster of equivalent data they drop it there. 5 . 5 . 4 A cl o se d s y s t em The present system is open because it has a constant stream of events feeding the info-agents as often as they require new data, and a loss of information when the arch-agents eliminate it after a period of inactivity. A closed system is achieved by cutting the stream of events coming from the Data space and keeping a constant amount of information inside the system to be managed by the agents. Characteristic of such a system is a closed feedback loop of interaction between agents that produces no loss of information (energy). The information that is passed to the arch-agents is returned to the info-agents later on, and reset in the cycle. Creating a closed system is not realistic in terms of the proposed scenario (or any other real world scenario), but can be useful to study emergent patterns of data organization and landscape behaviour. The data is continuously exchanged between infoagents and arch-agents reorganizing it in various ways. The initial settings can be observed over a longer period of time and consistently occurring behaviours noticed. This the only way of learning how the system behaves and from there one can use the most interesting/useful settings in the open system situation. 20 6 . C O N CL US I O NS In this project I present first of all an information architecture built as a system: - Composed of interconnected, organized components: the info-agents and arch-agents, - These components are affected from their existence in the system: the info-agents movement and the arch-agents adaptive behaviour, - It does something: it gives shape to data flow and usage, - It is of special interest: becomes an interactive interface to the self-organised collection of data. fig. 39 Sequence from a flock of Boids – Craig Reynolds The design of a self-organizing system is a new type of design, based on controlling the actions of the parts and not of the whole [Resnick, 1998, p.24]. The presented system follows some principles from self-organizing multi-agent systems producing an information landscape that behaves like a living system. Some aspects of this approach give an insight on a new paradigm of inhabiting Cyberspace bringing together scientific models and architectural intuitions. 6.1 THE CENTRALIZED MINDSET The Centralized Mindset is a term used by Mitchell Resnick to describe the approach of people to a problem when engaged in a design process. This project uses a decentralized approach designing the individual actions of the agents, but delegating the overall result to the interactions of multiple autonomous software agents [fig.65]. The agents have an autonomous character by the fact that they communicate through stimuli. They don’t instruct other agents how to act, only trigger a response in the others. The internal state of the receiver agent determines the action; the content of the stimulus determines quality of this action, like in neurons from the nervous system [Coveney, 1995, p.289]. The content is the data transferred between agents when they communicate. The quality of the action is a parameter that defines its scope, but not the type of action. An element of the system that might raise some questions about its role in terms of centralized mindset is the Black box. It is not conceived as a centralizing unit to manipulate the overall activity of the system. It doesn’t help agents to communicate beyond their local boundary, like a mother spaceship informing robots of every others success [Resnick, 1994, p.127]. And it doesn’t synchronize their actions in an explicit sequence towards some expected collective behaviour. Some agent models, like SWARM, have a centralized scheduler to manage these sequences of actions. The reason for it is that several autonomous behaviours don’t always define a single consistent state of the system [Burkhart, 1997]. And in scientific research on 21 simulation of real phenomena control over the result is important. The tendency is to “engineer” the emergent behaviour stealing a certain level of surprise from the system, but the system has to be biased towards very specific results. For my model of mobile agents in search for emergent behaviour the level of agent freedom is of utmost importance. Every agent sends events into the Black box and acts upon receiving an event. The processing of these events is natural and uncontrolled, only depending on the speed of processing the event stack, the natural stochasticity of the operating system. It produces unexpected behaviours, like the “air traffic” behaviour, that give a natural capricious feeling to the system. For the functional performance of the present system the use of tags is very important. It relates agents’ behaviour with the fig. 40 A sample of transformed space by a multiagent system, a fixed moment in time. fig. 41 Multitude of information on the WWW. Picture from RIOT browser by potatoland.org information content, allowing them to be selective when communicating with the neighbours [Holland, 1995, p.14] without a need of some central monitoring system. It ensures a certain segregation of information content without which the landscape would be partitioned in very small clusters of content in a confusing pattern. Still, the resulting patterns don’t follow any particular principle of data classification. But as McLuhan puts it, in our world with floods of information there has to be a shift into pattern recognition. One can no longer go through documents one by one to classify and order them. One has to quickly browse a sea of references and move to the areas where particular patterns are identified [fig.67]. The patterns resulting from the system don’t need to be meaningful under the traditional conceptions of data classification. New patterns emerge from the “natural” order of the system, and one just has to learn how to read them, through experience and practice. This system was modelled with a bottom-up approach without a concern about obtaining a particular collective behaviour. Nevertheless some top-down features could be introduced in order to improve its functionality. Increasing info-agent awareness of the global environment improves their efficiency in reaching a suitable area to leave the type of data they carry. This can be achieved by having a permanent record of the landscape status, which is consulted by info-agents before selecting fig. 42 ‘Websom’ – self-organizing maps for internet exploration, at two magnifications 22 a target. Giving info-agents a different behaviour according to the type of data would be another option. The info-source analyses incoming data and retrieves a tag, which translates into specific parameters for the agent’s behaviour. These differentiated agents produce differentiated patterns clearly identifiable and associated with a certain type of data. The only requirement is that incoming data carries a header with content information that the info-source knows how to translate. fig. 43 The information landscape as a living organism 6.2 THE LANDSCAPE METAPHOR fig. 45 Nurbs display mode from ‘Cichlid’ fig. 44 Information landscape with ‘Cartia’ This MAS uses the landscape metaphor to display the collected information on the virtual environment. This is a recurring metaphor in other information systems [figs.68, 69] and in the general display of quantifiable data or abstract concepts like search space of solutions [fig.70]. It is an advantage to use this metaphor to convey information because concepts like height, steepness or regularity have an understood symbolism and can be readily explored. Furthermore the landscape metaphor allows several viewing perspectives: plan view, a distant axonometric view or an immersive exploration of the landscape. Each 23 perspective offers a different reading of the data, and at any level the interpretation is straightforward. The landscape is something that the user comes across in many situations - from 2D maps to real life experience - thus is accustomed to interpret. Also in architectural design the landscape surface is a contemporary form of expression using information to shape these topological surfaces [fig.71]. The use of mobile agents in modelling these surfaces can translate any form of dynamic data into flows of temporal expression. Music, images or movies can be decomposed in an abstract data structure that generates expressive landscapes. The particular aspect of this landscape is that it changes obeying the time flow characteristic of the dynamics of the system. The system exhibits slow and fast dynamics [Holland, p.165-166]. There is a long time-scale structure for slow dynamics the arch-agents - and short time-scale elements for the fast dynamics - the info-agents [fig.72]. The arch-agents build the landscape for most of the user’s activity requiring slow dynamics that favour navigation, perception, understanding and recollection of the environment. They aren’t static but they are more stable: they don’t change so often, they don’t change so fast, and the changes are subtle. The info-agents have a more transient nature, moving and changing fig. 46 Diagram of the two time scales of the system character more often. They participate in a faster cycle of events, whose dynamics is more readable on short periods of time, and not necessary to remember in long term. The record of events is transferred to the arch-agents. In working with a medium that is a dynamic system the retrieval of objective results is difficult and requires long testing. The first reason is that the evolution of the MAS is mostly time dependent. It’s an open-ended process that cannot be accelerated. The second is that it presents an interactive experience that influences the end result. It is necessary to explore the virtual environment actively and in real time. But, taking it also as a functional medium, the potential of it in terms of usage has to be explored by changing parameters and doing experiments that sometimes clash with the initially established principles. These can bring new insights into how the system is working and which principles are really driving the desired dynamics. To play with such a phenomenon towards unexpected results one has to understand it fully, in both theoretical and practical terms. Yet much reward comes from a certain freedom of experimentation and the pleasure of surprises, obtained from an intuitive and uncompromised attitude. Engineering emergence is a current practice in scientific research on multi-agent systems and a very complicated (paradoxical?) challenge. Sometimes the urge to obtain the expected results introduces constrains that remove the organic nature of such systems and the necessary holistic mindset. The digital universe has its own ways and its own nature that has to be taken in consideration. The programming languages used to develop the system produced good results. The use of VRML brings the system beyond the prototype stage because it is a language that is used on the Internet, placing this architectural system in its natural context 24 of Cyberspace. In terms of geometries for further architectural space exploration it offers the possibility of using objects created in a variety of 3D modellers since many have a VRML exporter. The real-time navigation ensures an interactive spatial and visual experience to the user, which is very important in the terms of the proposed scenario. Although JavaScript isn’t the fastest language for processing the desired dynamics, it still allows a high degree of complexity. Since the rules for the control algorithms are simple it is able to handle with the designed behaviours. Only the amount of calculations becomes a problem, and after a longer period of execution the system starts to slow down. A natural step would be to translate the behaviours into Java classes that can be easily integrated in VRML and give a much higher performance. fig. 47 Stages of development of the system With this multi-agent system a step was given from one to many. Many objects, many types of objects and many media [Resnick, 1998, p.46]. Having many agents of the same type will produce repeatedly the same type of collective behaviour. The more agents you have, the faster the system will perform initially. If this is fine-tuned towards a desired result the system becomes efficient but less flexible. Having many of different types takes control from the system because the interactions are more complex, but on the other hand it can produce unexpected and useful collective behaviours. Further steps need to be given to start implementing such multi-agent systems in a multi-user and content rich medium like the Internet. First it can be enhanced by the introduction of new geometries that turn it more interesting visually and a richer architectural experience. But also the dynamics at the landscape level have to be enhanced: the utility of the system rises from the interaction between info-agents and info-landscape in self-organising the incoming information. That’s where the interface between the user and the available data is developed, thus the resulting patterns must be useful. Then the system has to be transferred to the real territory of information. No more a simulated system - as an open system it is able to start its eternal cycle like an ecosystem in cyberspace [fig.75]. The evolving virtual environment produces continuously new readings for the user to explore giving an updated and organized display of the information: it constitutes a living data space. Furthermore the multi-agent system is able to maintain a living architecture in cyberspace as an ecosystem, an open system. It transfers into the architectural object the dynamics of the context where it is inserted, making it adapt over time to the changing context, like a living organism. 25 9 . R EF ER E N C ES Allen, Stan [1997] ‘From object to field’ in Architecture after Geometry, p. 24, Architectural Design Profile n. 127, London, Academy Editions Benedikt, Michael [1991] ‘Cyberspace: some proposals’ in Cyberspace - First steps, p.119, Cambridge MA, MIT Press Burkhart, Roger [1997] ‘Schedules of activity in the Swarm simulation system’ - Position paper for OOPSLA ’97 Workshop on OO Behavioral Semantics, http://www.swarm.org/intropapers.html (Last accessed August 2000) Cook, Peter (editor) et al. [1999] Archigram, New York, Princeton Architectural Press Collins, Robert J. and Jefferson, David R. [1992] ‘AntFarm: Towards Simulated evolution’ in Artificial Life II – proceedings of the workshop on artificial life (Santa Fe Institute, Studies in the sciences of complexity, vol. X), p.579, Addison Wesley Publishing Company Coveney, Peter & Highfield, Roger [1995] Frontiers of complexity – the search for order in a chaotic world, London, Faber and Faber DECOI [1999] ‘Aegis Hyposurface – Autoplastic to Alloplastic’ in Hypersurface Architecture II, pp. 60-65, Architectural Design Profile n.141, London, Academy Editions Deutsch, David [1997] The fabric of reality, London, Penguin Books Egerton, P. A. & Hall, W. S. [1999] Computer Graphics – mathematical first steps, Prentice Hall Encyclopaedia Britannica, http://members.eb.com (last accessed August 2000) Frazer, John [1995a] An evolutionary architecture, London, Architectural Association Frazer, John [1995b] ‘The architectural relevance of cyberspace’ in Architects in Cyberspace, p.76, Architectural Design Profile n. 118, London, Academy Editions Frazer, John and Rastogi, Manit [1998] ‘The new canvas’ in Architects in Cyberspace II, p. 9, Architectural Design Profile n. 136, London, Academy Editions Frazer, John and Julia [1998] ‘The Groningen Experiment: Architecture as an artificial life form – Materialisation phase II’ in Architects in Cyberspace II, p. 13, op. cit 26 Fuller, R. Buckminster [1979] Synergetics folio, Philadelphia PA, R. Buckminster Fuller & Lim Chong Keat Girardin, Luc [1996] ‘Mapping the virtual geography of the World Wide Web’, in Fifth International World Wide Web conference, http://www.girardin.org/luc//cgv/www5/index.html (Last accessed August 2000) Gibson, William [1984] Neuromancer, London, Harper Collins Ho, Mae-Wan [1997] ‘The new age of the organism’ in New Science = New Architecture?, pp.44-51, Architectural Design Profile n.129, London, Academy Editions Hofstadter, Douglas R. [1980] Goedel, Escher, Bach – an eternal golden braid, London, Penguin Books Holland, John H. [1995] Hidden Order: how adaptation builds complexity, Reading MA, AddisonWesley Holland, John H. [1998] Emergence – From chaos to order, Oxford, Oxford University Press Holtzman, Steven H. [1994] Digital Mantras – the languages of abstract and virtual worlds, Cambridge MA, MIT Press Hraber, Peter T.; Jones, Terry and Forrest, Stephanie [1997] ‘The ecology of Echo’ in Artificial Life III – proceedings of the workshop on artificial life (Santa Fe Institute, Studies in the sciences of complexity, vol. X), pp. 165-190, Addison Wesley Publishing Company Hunt, Gillian [1998] ‘Architecture in the ‘cybernetic age’’, in Architects in Cyberspace II, p. 53, op. cit Kipnis, Jeff [1998] ‘(Architecture) After geometry – An anthology of mysteries: case notes to the mystery of the school of fish’ cyberspace’ in Architects in CyberspaceII, p.43, op.cit. Langton, Christopher G. [1989] ‘Artificial Life’ in Artificial Life – the proceedings of an interdisciplinary workshop on the synthesis and simulation of living systems (Santa Fe Institute, Studies in the sciences of complexity, vol. VI) p.1, Addison Wesley Publishing Company Levy, Steven [1992] Artificial Life, London, Penguin Books Maas, Winy [1999] Metacity / Datatown, Rotterdam, 010 Publishers 27 Marrin, Chris & Campbell, Bruce [1997] Teach yourself VRML 2 in 21 days, Indianapolis IN, Sams.net Maturana, Humberto [1997] ‘Metadesign’, INTECO, Instituto de Terapia Cognitiva http://www.inteco.cl/articulos/006/texto_ing.htm (Last accessed August 2000) McLuhan, Marshall [1964] Understanding media – the extensions of man, Cambridge MA, MIT Press McLuhan, Marshall [1967] The medium is the massage, San Francisco CA, HardWired Miller, James Grier [1995] Living Systems, Niwot COL, University Press of Colorado Minsky, Marvin [1988] The society of mind, New York, Simon & Shuster Mitchell, William J. [1996] City of bits – space, place, and the infobahn, Cambridge MA, MIT Press Novak, Marcos [1998] ‘Next Babylon, Soft Babylon – (trans)Architecture is an algorithm to play in’ in Architects in Cyberspace II, p. 21, op. cit Open Systems Group (editor) [1988] Systems behaviour, London, Paul Chapman Publishing Pesce, Mark D.; Kennard, Peter and Parisi, Anthony [1994] ‘Cyberspace’ in First International Conference on the World Wide Web, Geneva, Switzerland http://www.hyperreal.org/~mpesce/www.html (Last accessed August 2000) Pfenningstorff, Barbara [2000] ‘Inhabiting Information – The architecture of cognitive amplification’, http://www.bartlett.ucl.ac.uk/ve/Work9900/barbara/WebPage/EssayHtml/ EssayHtml.htm (Last accessed August 2000) Prigogine, Ilya [1984] Order out of chaos – man’s new dialogue with nature, London, Heinemann Rasmussen, Steen; Knusden, Carsten and Felberg, Rasmus, [1992] ‘Dynamics of programmable matter’ in Artificial Life II, pp.211-254, op. cit. Resnick, Mitchel [1989] ‘Lego, Logo and Life’ in Artificial Life op. cit., p.397, Addison Wesley Publishing Company Resnick, Mitchel [1994] Turtles, Termites and Traffic Jams – Explorations in massively parallel microworld’, Cambridge MA, MIT Press 28 Reynolds, Craig [1999] ‘Boids (Flocks, Herds, and Schools: a distributed behavioural model)’ http://www.red3d.com/cwr/boids/ (Last accessed August 2000) Saunders, Peter T. [1997] ‘Nonlinearity: what it is and why it matters’ in New Science = New Architecture?, pp.52-57, op.cit. Schmitt, Gerhard [1999] Information Architecture – Basis and Future of CAAD, Basel, Birkhauser Simon, Herbert [1981] The sciences of the artificial, Cambridge MA, MIT Press Taylor, Charles E. [1989] ‘RAM: Artificial Life for the exploration of complex biological systems’ in Artificial Life, p.275, op. cit Titman, Mark [1995] ‘Zip, Zap, Zoom – A Z-A of Cyberspace’ cyberspace’ in Architects in Cyberspace, p.48, op.cit. Toffoli, Tomaso and Margolus, Norman [1987] Cellular Automata Machines, Cambridge MA, MIT Press Tufte, Edward R. [1990] Envisioning information, Cheshire CON, Graphics Press Watt, Stuart [1997] ‘Artificial societies and psychological agents’, in British Telecom Technology Journal 14(4) Wiener, Norbert [1948] Cybernetics: or control and communication in the animal and the machine, Cambridge MA, MIT Press Wolfram, Stephen [1982] ‘Cellular automata as simple self-organizing systems’, Caltech preprint CALT-68-938 Wolfram Stephen [1988] ‘Complex systems theory’ in Emerging Syntheses in Science: Proceedings of the Founding Workshops of the Santa Fe Institute, pp. 183-189, AddisonWesley Yaeger, Larry [1994] ‘Computational genetics, physiology, metabolism, neural systems, learning, vision, and behaviour or Polyworld: Life in a new context’ in Artificial Life III, op. cit. 29 A P PE N DI X D : L O G OF T H E PR O J E C T D E VEL O P M EN T INITIAL TESTS Studying agent movement (from M. Lewis example) and interaction with the surface. Changed some parameters that control movement, producing different results. Here the turning radius is very small producing peeks. Introducing free random movement and several agents. Trying with spheres because they are non-directional shapes, and hide some frantic movement Scaling the surface to study the mechanism of communication between mobile agent and static surface. Testing a simple information display to control data from the scene. 30 First landscapes The systems geometry is set. The surface and the info-source are scaled to the final size. The agent’s position resulting from random movement is continuously communicated to the surface, thus changing it. The surface becomes a landscape, with the agent activity hidden. The two layers: architecture and information. An initial source is introduced. 31 Demonstration of the System Top view of the surface. Navigating the surface. Multiple mobile agents change the surface according to their movement. The centre grows more because they always return to the source. 32 A ge nt c o m m un ic at io n The surface is modulated by the info-agent when it reaches the target to drop data and sends a stimulus to the arch-agent. The result is an even growth in every part of the surface. A colour code is passed to the surface to more easily assess the communication between agents. Some observations are only possible by comparing the colour of the agents with the colour passed to the surface. And also the resulting “painting”. 33 Info-agents become selective To avoid the too densely chaotic pattern a different spacing of agents is tested. It is observed that agents replace the colour of arch-agents: They mix different types of data. A new movement is added: the agents select a new target when the first is not of a certain type. Here they move constantly. The selective behaviour is introduced. The info-agent only drops data in areas with the same type. 34 Testing with 40 info-agents Some tests to observe the system working with 40 info-agents. There is a correct communication and distribution of data. The limits of the grid become a boundary beyond which the agents cannot move. Avoids them running away when minimized, and taking too long to return. 35 Introducing the arch-agent neighbourhood The result is still dense and chaotic. The Moore neighbourhood is introduced to promote the concentration of data of the same type. A tag is passed to archagent neighbours giving them the same type. The colours of the landscape spread over larger areas. 36 For the user interface To facilitate testing with different info-agent numbers a script generates them. It is only necessary to change a number to specify how many agents are supposed to run, and all the code is dynamically created. This is a test with 100 agents. After knowing that the communication is working, the colours are removed from the surface. Only the height changes. Introduction of links at active arch-agents to allow the user to identify the content of mountains and link to data content. 37 The links don’t face the user. Reading becomes more difficult. The links are oriented at the moment they are generated, but don’t dynamically update their orientation. The user can access the content of info-agents. A name appears on the top-right corner of the screen. Info-agents in empty state towards the source give a message “collecting information”. By clicking on the agent with the mouse a window pop’s up with a file. This can be of several formats. Here it is an image. 38 A more subtle augmented reality style information display is introduced. Links are placed under the surface to allow a reading of data organization when observing the info-source. The links always face the user. 39 But they don’t update when the arch-agent changes its height. 40 Links are now fully dynamic, facing the user and updating their height. The information is gathered in larger “hills” to avoid greater fragmentation of the surface. Viewpoints for following an infoagent and an arch-agent are created. They give a greater understanding of their roles. The info-agents have reverse behaviour resetting areas after some time. 41