Contact information. Tony White, Associate Professor

Transcription

1 Contact information Tony White, Associate Professor Office: Hertzberg 5354 Tel: x2208 Fax: Web:

2 Announcement Amorphous Computing (1 lecture) Per Bak (1-2 lecture) How Nature Works: the science of self-organized criticality Camazine et al (1 lecture) Chapter 8, Self Organization in Biological Systems Pattern Formation in Slime Molds and Bacteria Lectures required for 1 st, 2 nd and possibly 3 rd classes in February (4 th, 7 th and 11 th )

3 Announcement Amorphous Computing (1 lecture) Resnick (1 lecture) Turtles, Termites and Traffic Jams Starlogo Lectures required for 1 st, 2 nd and possibly 3 rd classes in February (4 th, 7 th and 11 th )

4 Summary Discussed Agent as a black box Deliberative vs reactive architecture Symbolic nature of models Communication Special structures: blackboards and protocols Centralized Weaknesses Brittle Don t scale

5 Swarm Intelligence a.k.a. Self-Organizing Systems Lecture 3

6 What is Swarm Intelligence? Swarm Intelligence is a property of systems of non-intelligent robots exhibiting collectively intelligent behavior. [Beni] Characteristics of a swarm: distributed, no central control or data source; no (explicit) model of the environment; perception of environment, i.e. sensing; ability to change environment.

7 What is Swarm Intelligence? Swarm systems are examples of behaviorbased systems exhibiting: multiple lower level competences; situated in environment; limited time to act; autonomous with no explicit control provided; problem solving is emergent behavior; strong emphasis on reaction and adaptation;

8 What about Software Agents Differences I would say! No global controller Immersed in environment No symbolic view of the world Environment is memory Communication is not directed No KQML, dialogs or contract net protocols

9 EE141:Swarm Intelligence Lecturers: Alcherio Martinoli Rod Goodman Owen Holland Adam Hayes TAs: Adam Hayes William Agassounon Kjerstin Easton Philip Tsao Joseph Chen Amit Kenjale

10 Swarm Intelligence Swarm Intelligence: Any attempt to design algorithms or distributed problem-solving devices inspired by the collective behavior of social insect colonies and other animal societies. [Bonabeau, Dorigo, and Theraulaz, 1999] Minimalist but fully autonomous individuals Fully distributed control Exploitation of robot-robot and robot-environment interactions Exploitation of explicit or implicit (stigmergic) communication Scalability (from a few up to thousands individuals) Enhanced robustness through redundancy and minimalist design of the individuals

11 Biological Motivation Biological Inspiration from: social insects (ants, bees, termites) flocks of birds, herds of mammals, schools of fish, packs of wolves, pedestrians, traffic. Colonies of social insects can achieve flexible, reliable, intelligent, complex system level performance from insect elements which are stereotyped, unreliable, unintelligent, and simple. Insects follow simple rules, use simple local communication (scent trails, sound, touch) with low computational demands. Global structure (e.g. nest) reliably emerges from the unreliable actions of many.

12 All rights reserved

13 All rights reserved

14 All rights reserved

15 Guy Theraulaz

16 All rights reserved

17 Pascal Goetgheluck

18 Human Swarms All rights reserved

19 Insect Societies A natural model of distributed problem solving Collective systems capable of accomplishing difficult tasks, in dynamic and varied environments, without any external guidance or control and with no central coordination Achieving a collective performance which could not normally be achieved by any individual acting alone Constituting a natural model particularly suited to distributed problem solving Many studies have taken inspiration from the mode of operation of social insects to solve various problems in the artificial domain

20 Insect Societies Individual simplicity and collective complexity The behavioural repertoire of the insects is limited their cognitive systems are not sufficiently powerful to allow a single individual with access to all the necessary information about the state of the colony to guarantee the appropriate division of labour and the advantageous progress of the colony the colony as a whole is the seat of a stable and self-regulated organisation of individual behaviour which adapts itself very easily to the unpredictable characteristics of the environment within which it evolved

21 Self-organisation Systems of collective decision-making Insect societies have developed systems of collective decisionmaking operating without symbolic representations, exploiting the physical constraints of the environment in which they evolved, and using communications between individuals, either directly when in contact, or indirectly (stigmergy) using the environment as a channel of communication Through these direct and indirect interactions, the society selforganises and, faced with a problem finds a solution with a complexity far greater than that of the insects of which it is composed

22 Collective or Swarm Intelligence Some questions... How do insect societies manage to perform difficult tasks, in dynamic and varied environments, without any external guidance or control, and with no central coordination? How can a large number of entities with only partial information about their environment solve problems? How can collective cognitive capacities emerge from individuals with limited cognitive capacities?

23 Intelligence What do we mean by intelligence? Ability to solve problems in some abstract or real domain Ability to produce behaviour appropriate to a situation What is it that has intelligence? A natural system of one or more agents An artificial system of one or more agents (computational entities or robots) At opposite ends of the scale: Human intelligence vs. insect intelligence How useful is each of these as a model for intelligent artificial systems?

24 Intelligence Individual human intelligence Highly capable, extremely flexible Consciously reason about the problem, seeking new information where necessary, and generate and execute a plan. Individual artificial intelligence Capable in niche areas, inflexible Apply rules of logic and reason to an abstract representation of the problem situation, seeking new information where necessary, and generate and execute a plan. Conventional robot Incapable compared with humans, inflexible. Build a representation of the problem situation, and apply rules of logic and reason to it, seeking new information where necessary, and generate and execute a plan.

25 Individual insect intelligence Intelligence Extremely capable and flexible within niche (specialist) situations, generally incapable outside them. Cued to environment incapable outside Specialised behaviours triggered by specialised sensing; chaining of behaviours by internal or external cues; suppression of some behaviours by others. Insect level (behaviour based) robots MIT 85 insect lab Rod Brooks AI lab Excellent at low level tasks in particular environments, often flexible, easy to build, often robust to component failure. Mimic individual insect intelligence - Specialised behaviours triggered by specialised sensing; chaining of behaviours by internal or external cues; suppression of some behaviours by others.

26 Sheepdog behavior based robot Richard Vaughan

27 Multiple vs. Individual Systems Multiplication of effort, resources etc. Distributed sensing Distributed action Division of labour Specialisation Extended time scales Redundancy etc BUT The BIG QUESTION How can individual efforts be coordinated to achieve the common goal?

28 Multiple vs. Individual Systems Multiple humans The most powerful systems we know, but vulnerable to high-level failures. Dominated by communications. centralized hierarchical: information up, decisions down highest level has global view of situation specialists explicit task allocation and reallocation cognitive level communications Etc Multiple AI systems Modelled on humans; too soon to judge. Multiple conventional robots Almost an empty category; a few toy systems, mostly simulated.

29 Multiple Insects Colonies of social insects (ants, termites, some wasps, some bees) are spectacularly capable. Complex lifestyles Flexible and robust The capabilities of the colony transcend those of the individual insects The individual insects appear to be no more capable than solitary insects, but many of their behaviours are directed towards affecting the behaviour of others. How is this intelligence achieved? Self organisation: local interactions between insects, and between insects and the environment, produce emergent patterns and configurations which solve the problem. Specialisation: By morphology (soldiers etc castes) By behaviour (identical ants do different tasks stay with that task over time. E.g. brood, or midden stay with particular behaviour By age older more dangerous )

30 Swarm-Based AI Ant Colony Optimisation extremely successful at a variety of very difficult combinatorial optimisation problems some of the best solutions known to some problems Ant Based Control of telecommunications systems exceptionally flexible solutions best solution for some problems

31 Swarm-Based Collective Robotics Multiple robots offer significant advantages over single robots: Simultaneous sensing and action in multiple places Task dependent reconfigurability Enhanced system performance through work division Robustness through redundancy Task enabling if the task could not be solved by an individual we can build systems of behaviour based robots with adequate capabilities we can demonstrate the use of self-organisation by these systems in task achievement Gilles Caprari

32 Multiple vs. Single Robotics Solutions +Task enabling. +Distributed sensing and action. +Enhanced system performance by work division. - Interference. - Increased uncertainty and dynamics. - Overall system cost. Autonomy at the... Energetic level. Sensory, motor, and computational level. Decisional level.

33

34 A Few Concepts Locality and Globality based on physical proximity. E.g. local and global communication. E.g. local and global information. Flexibility is the capacity of a society to change its collective behavior [Gordon92]. Plasticity is defined as the capability of individuals to adapt their control parameters. Adaptation implies not only the capacity for change, but the additional requirement that this change represents an improvement in fit [Belew96]. Centralized and Decentralized Team Control Centralized: central unit coordinates the group decisional processes. Decentralized: no central coordination. Hierarchical and Distributed Team Control Hierarchical control: locally centralized. Distributed control: each teammate has full decisional power. An intelligent individual is able to... act in its environment so that a viability condition is always satisfied. maintain its identity (in a broad sense). A team is provided with collective intelligence if the viability of the team is required in order to achieve the viability of the individual.

35 Robots, techniques, and tools Probabilistic Modeling Simulation Point Embodied Real Robots Moorebot Kephera Alice Tools Overhead vision system for tracking, monitoring KPS Powered floor for Kephera Charging docks for Moorebot

36 MooreBots

37 MooreBots We have built 15 robots based on an O. Holland design. We will use these to test our theories, and to perform experiments on collective robotics. The Robots Feature: Wireless LAN communication to each other, base station, and Internet Linux operating system PC 104 (386 to Pentium) Architecture PCMCIA slots for GPS, frame grabber, etc. 10 dia size 2 m.p.h., 720 degrees/sec 4 miles range, 2 hours endurance odometry to 0.1% accuracy wheeled or tracked chassis

38 Khepera Miniature Robot actuators sensors 68HC11 microcontroller (slave) microcontroller (master) batteries 5.5 cm

39 Alice Micro-Robot Developed by G. Caprari, Autonomous System Lab, EPFL, Switzerland We are working with the developers to put nose chips on the robot Main Features: Modular Dimensions: 22mm x 20mm x 19mm Max Speed: 35 mm/s Power <10mW Power autonomy up to 10 hours. 4 proximity sensors Local IR robot-robot communications Low power radio comms robot-robot and robot-host PC. Range 10m. PIC 16F84 with 1Kb Flash memory

40 Simulation Tools Non-embodied simulator (point simulator) Multiple-robot. No interference: robots simulated as points Acceleration ratio: up to thousands times than the same real robot experiment. Embodied simulator (Webots) Multiple-robot. Khepera, Alice, and soon Moorebots (customizable robot chassis, actuators, and sensors). Acceleration ratio: up to hundred times than the same real robot experiment. Smooth transition between simulator and real robots (API translator).

41 Some Further Tools... KPS: laser-based scanning, fully scalable (perfomances independent from group size), ±5 4 m, Hz Monitoring collective experiments with ceiling camera _ + + External power supply from a special electrical floor: position and orientation independence, depending on the configuration, open-end experiments

42 Experimental Levels Simulation Analytical probabilistic models Numerical probabilistic models Non-embodied simulations Embodied, sensor-based simulations Real robots Partial virtual environments Desktop/laboratory environments Real environments Abstraction Reality

43 Probabilistic Modelling and Collective Understanding and prediction of collective team performance based on single-robot features. Prediction of large swarm of robots. System optimization (number of robots, control parameters, body and actuator morphology, sensor and communication range and type, ). System Optimization Embodied simulator Real robots Probabilistic model

44 Collective Robotics/SI Applications

45 Swarms - Artificial

46 Movies - ANTZ "If you have a film about ants, it would be disappointing to never have a shot with more than five ants. You want 5,000--or maybe 50,000 ants to convey the sense of a colony to the audience," said Kirk. "With our crowd system, the directors and animators have a great amount of creative control, yet there's a tremendous labor savings. They're able to define the way a crowd will look, move, and behave, and then let the computers calculate the specific motion for each ant in the crowd."

47 Crowd Control

48 Khepera Stick-Pulling Experiments IDSIA

49 Beckers et al. Aggregation Experiments Biological inspiration from social insect aggregation processes (e.g. clustering of dead ant corpses) Probabilistic model (cluster modifying probabilities) Real robots

50 IDSIA Khepera Aggregation Experiments

51 Distributed Exploration:Robots Collaborate to find an IR Beacon Each robot has itself an omni-directional beacon which is used to signal when the robot sees the homing beacon, including when it arrives at the homing beacon. In collaboration mode the omni-directional beacons are enabled. In non-collaboration mode they are disabled. The task is completed when all N robots arrive at the beacon.

52 Collective Advantage Non-collaborative Non-collaborative Collaborative Collaborative Non-collaborative Collaborative

53 Traffic

54 Distributed Plume Tracing

55 History Swarm Intelligence first used by Beni, Hackwood and Wang in context of cellular robotics. Simple agents occupy 1 or 2 D environments. Self-organize through nearest neighbour interactions. Collections of simple agents or automata to solve problems on graphs or lattices in work of Butrimenko, Tsetlin, Stefanyuk and Rabin. Rabin introduced moving automata to solve problems on graphs or lattices by interacting with the consequences of their previous actions. Tsetlin identified the important characteristics of biologically-inspired automata that make the swarm-based approach potentially powerful: Randomness Decentralization Indirect interactions among agents and self-organization

56 History Butrimenko applied ideas to control telecommunications networks Stefanyuk applied ideas to cooperation between radio stations Lecture on Tsetlin automata would be worthwhile Nice talk or Project here

57 Emphasis Design of adaptive, decentralized, flexible and robust artificial systems capable of solving problems. Inspiration is social insects.

58 How to proceed? Have to understand the mechanisms that generate collective behaviour in insects. Modelling is key here. Different from designing an artificial system modelling tries to uncover what happens in the natural system. Should reproduce features of the natural system and be consistent with what is known about it: Parameters cannot be arbitrary Mechanisms must be plausible Model should be predictive: Predictions should be testable Variables and parameters should be experimentally verifiable

59 Engineering Does not have to be concerned with biological plausibility: Efficiency Flexibility Robustness Cost Natural selection uses survival of the fittest Essentially a tinkering process Social insects are remarkably successful

60 The use of Metaphor Neural Networks Abstraction of actual brain organization Genetic Algorithms Abstraction of basic evolutionary process Basic principles are present, but detail is unimportant. Ultimately, good problem-solving device does not have to be biologically relevant.

61 Self-Organization in Insects Self organization is a set of dynamical mechanisms whereby structures appear at the global level from interactions among its lowerlevel components. Rules specifying the interactions among system constituent units are executed on the basis of local information, without reference to the global pattern. Global pattern is emergent property of system, rather than a property imposed upon system by external controlling influence.

62 Graphically System System Controller

63 Self-organization (SO) in Social Insects SO is a set of dynamical mechanisms whereby structures appear at the global level of a system from interactions among its lower-level components. Arises as a consequence of statiotemporally organized networks of pheromone trails (for example). Is often inefficient, non-optimal Non-greedy search Agents (ants) can get lost

64 Self-organization in Social Insects Self organization relies on 4 ingredients: Positive feedback Preferential food source exploitation Negative feedback Used for stabilization: saturation, exhaustion or competition; e.g. Lotka-Volterra dynamics Amplification of fluctuations Random walks, errors, random task switching Often crucial, allows discovery of new solutions to occur Multiple interactions Signals from one individual have to be seen by others Environment is key element in system Density and signal strength are key

65 Important: Agent memory is NOT a requirement for patterns to develop. Memory is collective it s stored in the networks/lattices on which the agents operate.

66 Figure 1.9 From Swarm Intelligence Bonabeau et al.

67 Figure 1.10 From Swarm Intelligence Bonabeau et al.

68 SO Properties Statiotemporal structures develop in an initially homogeneous medium For example, characteristic well-organized pattern develops in honey bee combs Pattern consists of concentric regions: Brood area Rim of pollen Large peripheral region of honey

69 Figure 1.11 From Swarm Intelligence Bonabeau et al.

70 Bee Colonies 25,000 females Few thousand dones Single queen Also: Immature brood: eggs, larvae and pupae Honey and pollen ~ hexagonal 100,000 cells Temperature maintained degrees C Gradient across comb cannot explain pattern formation Isn t a colouring pattern mechanism either

71 Bee s the model Assumptions based upon experiment: Queen moves randomly over combs and lays ( per day) most eggs in neighborhood of cells occupied. Eggs remain for 21 days Honey and pollen are deposited randomly in selected cells (there s a pattern). Proximity of food a factor. 4x more honey than pollen is brought back to hive than pollen Typical removal input rations for honey and pollen at 0.6 and 0.95, respectively Removal of honey and pollen is proportional to the number of surrounding cells containing the brood.

72 Bees are amazing! Colony collects 60kg of honey 40mg per nectar load: ~ 3,000,000 trips! Pollen load is 15mg: ~1,300,000 trips! 25 nectar trips required to fill a cell 15 pollen trips required to fill a pollen cell

73 Where do they deposit material? Pollen foragers select a cell to deposit load Honey foragers regurgitate load to food storage bees, which select cell Pattern of deposits independent of whether brood is present or not: food often deposited within the brood area

74 Chapter 16 page 317 Self Organization in Biological Systems: Camazine et al.

75 Chapter 16 page 317 ~ Exponential distribution Self Organization in Biological Systems: Camazine et al.

76 Self Organization in Biological Systems: Camazine et al. Removal of pollen / honey

77 Conclusion from Experiments Honey and pollen are preferentially removed from cells near brood Magnitude is 10x Interacting processes of deposit and removal lead to characteristic brood pattern Queen has to return to brood area to discover cells emptied of food and brood Food is deposited most readily at boundary of brood and is consumed there too changes rapidly Cells reserved for brood are infrequently turned over Interface region emerges automatically