Amit Shabtay Zinovi Rabinovich Supervised by Jeffrey S
Amit Shabtay Zinovi Rabinovich Supervised by: Jeffrey S. Rosenschein In collaboration with: 1
Parasites- Paradigm Motivation • Our paradigm employs a special kind of agent (called “Behaviosite”) that manipulates the behavior of other agents. • Affecting the behavior of several agents in a distributed manner will facilitate altered performance of the entire system. • By definition, the behaviosite is not necessary for the normal conduct of the system, thus termed a kind of “parasite”. 2
Lecture Layout 1. Parasites in biological context and in computer science 2. Formalization of the Behaviosite Paradigm 3. 1. Presenting the paradigm in the El Farol problem and Behaviosite paradigm and floys Discussion and Future work. 3
Parasite Concept in Biology & CS 4
1 Parasite Concept Parasite In Nature • A parasite is an organism that lives inside or outside the living tissue of a host organism at the expense of it. • The biological interaction between the host and the parasite is called parasitism. The parasite usually harms the host, but not necessarily. • It can have a complex life cycle. • They may help the host, as in the case of bees. 5
1 Parasite Concept Parasites in Computer Science • Parasites appear in three forms in CS: – As an observed phenomena in evolution • Tierra Virtual World (Thomas Ray 1992) – As helpers in genetic algorithms using co-evolution. • Co-evolving parasites improving the sorting problem (Hillis WD. 1990 and many more examples) – As malware in the electronic world. • Parasite is a known concept: Computer viruses, Worms, Trojan Horses as parasites (R. J Bagnall). • Viruses today are more focused and interested in quietly stealing our data and control over the computer than just crashing it (Meet the Sonic Worm, Zone Alarm 2000) 6
Behaviosite Formalization 7
2 Behaviosite Formalization Behaviosites Formalization I • Behaviosites act as a society of special agents within a system composed a society of agents and environment A distributed solution to issues raised in a distributed environment • The behaviosite is an additional property/information added to the system (and not the agent). • Behaviosites must be beneficial to the system in some sense, not necessarily in regards to the initial purpose of the system. 8
2 Behaviosite Formalization Behaviosites Formalization II • Basically, behaviosites are designed in two levels: infection strategy and manipulation strategy. – Infection strategy: finding the best host to infect at the current time step and how to move between agents. – Manipulation strategy: possible options for the behaviosite to manipulate the behavior of the infected agent. One may also include “behaviosite ecology”- where do they come from? 9
2 Behaviosite Formalization Behaviosites Formalization III • Benefiting the system • Deep system knowledge • Use existing capabilities • Small numbers • Mobility between hosts 10
2 Behaviosite Formalization External vs. Internal Behaviosites • Behaviosites can alter the input or output of the agent vis-ávis the environment (external behaviosites) or using an internal hook (internal behaviosites). External Internal • An agent designer can have an incentive to create such a hook, if it is required of him, or if it can be guarantied that the overall performance of the agent will not degrade because of it. 11
2 Behaviosite Formalization Behaviosites Optional Traits • Hidden vs. Apparent infection. There are some settings in which the sheer knowledge that an agent is infected, is sufficient for the behavior manipulation. • Behaviosite communication. Behaviosites may communicate within an infected host or across hosts to form some kind of an inner network. 12
The El Farol Problem El Farol 13
3 The El Farol Problem • The El Farol problem is an example of a distributed system (Brian Arthur 1994), first suggested as a Congestion Problem in economics. • All agents want to go to a bar called “El Farol”, but it has a limited (comfortable) capacity. Attended and undercrowded Did not attend Attended and overcrowded • With no option for communication or collusion, an agent must learn the behavior of other agents en-masse, in order to reach a decision. 14
3 The El Farol Problem Parasitized El Farol Problem • The system reaches an equilibrium around the capacity, where every agent has a unique, simple learning decision algorithm. • However, personal and social utilities are suboptimal. • We show that using behaviosites with simple infection and manipulation strategies, both utility and social fairness improve, overcoming learning ability of agents. 15
3 The El Farol Problem Parasitized El Farol Problem • Infection strategy: infect all, infect attending, infect when overcrowded. • Manipulation strategy: lower the believed capacity of the infected agent (50 40, 60 40, 80 60). 16
3 The El Farol Problem Mean Attendance and Social Utility • Infect all had the most severe effect on attendance, while infect when overcrowded had the least effect. • Attendance for capacity of 60 • Utility for infect attending: 80 Overcrowded 60 Attending 50 All 17
3 The El Farol Problem Simulation Social Fairness • Formula for social fairness according to attendance: • For capacity of 60: Attending Overcrowded All 50% 18
Controlling a Swarm of Floys 19
3 The Floys Problem Controlling a Swarm of Floys • Controlling a swarm has received much attention (UGV, computer graphics) • Reynolds (1987) showed that it is possible to create a swarm behavior using three rules: – Separation – Cohesion – Alingment Rome 20
3 The Floys Problem A Swarm of Floys 21
3 The Floys Problem Controlling a Swarm of Floys • Infection Strategy: Jump to an uninfected floy within sight. • Manipulation Strategy: Make the floy move two “turn units” toward the goal point. If in vicinity of goal, switch to next goal. 22
3 The Floys Problem Tasks for Behaviosites • Keep swarm in one place • Move swarm between check points (rectangle, circle) • Move between equilibrium points 23
3 The Floys Problem Parasitized Swarm Simulation • It takes only 5% infection rate for achieving control Number of drawn rectangles Distance from true path 24
3 The Floys Problem Parasitized Swarm Simulation • Can create a movement of the swarm along a path • Robust to malfunctioning, ill-functioning, or destroyed behaviosites • Behaviosites are endemic, thus protected by the swarm from external harm • Few can control many • Behaviosites can move to the most effective position at a given time without disturbing the swarm (unlike herdsman). • All tasks were accomplished using only one infection and manipulation strategy, and one type of simple behaviosite. 25
Discussion & Future Work 26
4 Discussion & Future work Discussion • The core of the Behaviosite Paradigm is creating a distributed behavioral changes in a small number of agents using infection and manipulation strategies, to achieve a global effect. • We described the Parasitized El Farol Problem, and a method for controlling a swarm • Behaviosites are not a type of “lie” in the system, since they cannot be disregarded or overcome. 27
4 Discussion & Future work Future Work- Appetisers • Use behaviosites as an information propagation mechanism in array of sensors • Use behaviosites in a congestion problem like traffic routing (packet routing) 28
4 Discussion & Future work Future Work- Appetisers • Turn floys to boids and deal with obstacle avoidance • Automatic story generation 29
4 Discussion & Future work Future Work- Ant Foraging • Using behaviosites in a colony of ants foraging when food sources suddenly appear B Nest Food source A Infection Strategy? Manipulation Strategy? Ecology? 30
4 Discussion & Future work Future Work- Ant Foraging • Using behaviosites in a colony of ants foraging mutually exclusive appearing/disappearing food sources Nest A B Infection Strategy? Food source Manipulation Strategy? Ecology? 31
4 Discussion & Future work Future Work- Ant Foraging • Final stage- food sources appear and disappear randomly. Infection Strategy? Manipulation Strategy? Ecology? Combination of Behaviosites? 32
33
- Slides: 33