Loading in 5 sec....

Fast Computation of Population Protocols with a LeaderPowerPoint Presentation

Fast Computation of Population Protocols with a Leader

- By
**azure** - Follow User

- 81 Views
- Uploaded on

Download Presentation
## PowerPoint Slideshow about ' Fast Computation of Population Protocols with a Leader' - azure

**An Image/Link below is provided (as is) to download presentation**

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.

- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript

### Fast Computation of Population Protocols with a Leader

Dana Angluin(Yale), James Aspnes(Yale), David Eisenstat(Princeton)

The trend

- Centralized systems
- Distributed Systems
- WSN and mobile devices and RFID
- Smart molecules?

When sensors “meet”

Before C(u)=a, C(v)=b

After C(u)= a’, C(v)=b’

u

v

a, b ➝ a’, b’

initiator

responder

Asymmetric interaction

configuration

configuration

When sensors “meet”

Sensors are (usually) anonymous

Initial state X = initial state of all the sensors

Final state Y = final state of all the sensors

Output Z = f(Y)

u

v

initiator

responder

(A computation) given X, computes Y (or Z)

-- Interactions are random

-- Execution is a sequence of configurations (Non-deterministic)

-- No sensor has global knowledge

-- Protocols are (generally) non-terminating

More on population protocol

(Fair execution)

C➝ C’ and C occurs infinitely often ⇒ C’ occurs infinitely often

(Stable computation of a predicate P(X))

Every fair execution converges to a configuration that reflects the

correct value of P

Epidemic

(One way epidemic)

State Space {0,1}: 1 = infected, 0 = susceptible

The protocol: (x,y) ➝ (x, max(x,y))

(Question) Starting with a single infected agent, how many

interactions are needed to infect every agent?

Epidemic

Lemma 1. Let T(k) be number of interactions before a one-way

epidemic starting with a single infected agent infects k agents.

For any fixed ε > 0 and c > 0, there exist positive constants c1 and

c2, such that for sufficiently large n and anyk >nε,

c1.n lnk ≤ T(k) ≤ c2.n lnk

with probability at least (1 − n−c)

Epidemic

Proof hint. Uses instances of the “Coupon Collectors problem”

and then uses Chernoff bounds …

Phase clock

(Motivation) How can a leader figure out when the epidemic is likely to have finished? A leader can only count the number of its local Interactions before moving to the next phase.

(Overview of the result) The phase clock protocol suggests a way for the leader to count off θ(n.logn) local interactions, in order that it outlasts the completion of an epidemic with high probability.

Phase clock protocol

Clock phase ∈ {0,1,2,…,m-1}

Let b = any agent, x, y are clock phases

(x, b),(y, follower) ➝ (x, b),(max* (x, y), follower)

(x, b), (x, leader) ➝ (x, b), (x + 1 mod m, leader)

(x, b), (y, leader) ➝ (x, b), (y, leader) [y ≠ x]

*A follower in phase x copies the state from any initiator with phase in the range [x+1, x+m/2] A round consists of m phases 0..m-1. Successive rounds should be θ(n.logn) apart with high probability

Phase clock protocol

Lemma 2. Let phase istart at interaction t. Then there is a constant a such that for sufficiently large n, phase (i + 1) starts before interaction t + a.nlnnwith probability at most n−1/2.

Phase clock protocol

Theorem 1. For any fixed c, d > 0, there exists a constantm such that, for all sufficiently large n, the finite-state phase clock with parameter m, starting from an initial state consisting of one leader in phase 0 and n−1 followers in phase m−1, completes nc rounds of m phases each, where the minimum number of interactions in any of the nc rounds is at least d.nlnnwith probability at least 1 − n−c.

Duplication

A duplication protocol has state space {(1, 1), (0, 1), (0, 0)} and transition rules:

(1, 1), (0, 0) ➝ (0, 1), (0, 1)

(0, 0), (1, 1) ➝ (0, 1), (0, 1)

A duplication protocol starting with a “active” agents in state (1, 1) and the rest in the null state (0, 0) converges to 2a “inactive” agents in state (0, 1), provided 2a < n (otherwise it converges to a population of mixed active and inactive agents with no agents left in the null state).

Duplication

When the initial number of active agents a is close to n/2 , duplication may take as much asθ(n2) interactions to converge, as the last few active agents wait longer to encounter the last few null agents. But for smaller values of a the protocol converges more quickly.

Duplication

When the initial number of active agents a is close to n/2, duplication may take as much as θ(n2) interactions to converge, as the last few active agents wait to encounter the last few null agents. But for smaller values of a the protocol converges more quickly.

Lemma 3. Let (2a+b) ≤ n/2. The probability that a duplication protocol starting with a active agents andbinactive agents has not converged after (2c+1).n lnn interactions is at most n−c.

Cancellation

A cancellation protocol has states {(0, 0), (1, 0), (0, 1)} and transition rules:

(1, 0), (0, 1) ➝ (0, 0), (0, 0)

(0, 1), (1, 0) ➝ (0, 0), (0, 0)

Cancellation

The cancellation maintains the invariant that the number of 1 tokens in the left-hand position minus the number of 1 tokens in the right-hand position is fixed. It converges when only (1, 0) and (0, 0) or only (0, 1) and (0, 0) agents remain.

As with duplication, the number of interactions to converge when (1, 0) and (0, 1) are nearly equally balanced can be as many as θ(n2), since we must wait in the end for the last few survivors to find each other

Food for thought

-- New computational model that may be naturally supported in

certain settings. The authors propose “computation by epidemic”

that can mimic many operations of a conventional register machine.

-- Are there ways to speed up some of these operations?

-- Solutions for new problems on this model

-- Can we eliminate the leader without drastically raising the cost?

-- New models of agent interaction to reflect the physical effects of the

spatial dispersion of agents.

Download Presentation

Connecting to Server..