1 / 48

Mean Field Methods for Computer and Communication Systems

Mean Field Methods for Computer and Communication Systems. Jean-Yves Le Boudec EPFL ACCESS Distinguished Lecture Series, Stockholm, May 28, 2012. Contents. Mean Field Interaction Model Convergence to ODE Finite Horizon: Fast Simulation and Decoupling assumption

bella
Download Presentation

Mean Field Methods for Computer and Communication Systems

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. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Mean Field Methods for Computer and Communication Systems Jean-Yves Le Boudec EPFL ACCESS Distinguished Lecture Series, Stockholm, May 28, 2012 1

  2. Contents • Mean Field Interaction Model • Convergence to ODE • Finite Horizon: Fast Simulation and Decouplingassumption • Infinite Horizon: Fixed Point Method and Decouplingassumption 2

  3. 1 Mean Field Interaction Model 3

  4. Mean Field • A model introduced in Physics • interaction between particles is via distribution of states of all particle • An approximation method for a large collection of particles • assumes independence in the master equation • Why do we care in information and communication systems ? • Model interaction of manyobjects: • Distributedsystems, communication protocols, gametheory, self-organizedsystems 4

  5. Mean Field Interaction Model • Time is discrete • N objects, N large • Object n has state Xn(t) • (XN1(t), …, XNN(t)) is Markov • Objects are observable onlythroughtheir state • “Occupancy measure”MN(t) = distribution of object states at time t 5

  6. Mean Field Interaction Model • Time is discrete • N objects, N large • Object n has state Xn(t) • (XN1(t), …, XNN(t)) is Markov • Objects are observable onlythroughtheir state • “Occupancy measure”MN(t) = distribution of object states at time t • Theorem[Gast (2011)] MN(t) is Markov • Called “Mean Field Interaction Models” in the Performance Evaluation community[McDonald(2007), Benaïm and Le Boudec(2008)] 6

  7. A Few Examples Where Applied Never again ! E.L. 7

  8. Example: 2-Step Malware Mobile nodes are either `S’ Susceptible `D’ Dormant `A’ Active Time is discrete Nodes meet pairwise (bluetooth) One interaction per time slot, I(N) = 1/N; mean field limit is an ODE State space is finite = {`S’ , `A’ ,`D’} Occupancy measure isM(t) = (S(t), D(t), A(t)) with S(t)+ D(t) + A(t) =1 S(t) = proportion of nodes in state `S’[Benaïm and Le Boudec(2008)] Possible interactions: Recovery D -> S Mutual upgrade D + D -> A + A Infection by active D + A -> A + A Recovery A -> S Recruitment by Dormant S + D -> D + D Direct infection S -> D Direct infection S -> A 8

  9. Simulation Runs, N=1000 nodes Node 1 State = D State = A State = S Node 2 Node 3 D(t) Proportion of nodes In state i=1 A(t) Proportion of nodes In state i=2 9

  10. Sample Runs with N = 1000 10

  11. Example: TCP and ECN ECN Feedback q(R(t)) 1 ECN router n queue length R(t) N N connections • [Tinnakornsrisuphap and Makowski(2003)]At, every time step, all connections update their state: I(N)=1 • Time isdiscrete, meanfieldlimitisalso in discrete time (iteratedmap) • Similarexamples: HTTP Metastability[Baccelli et al.(2004)Baccelli, Lelarge, and McDonald]Reputation System [Le Boudec et al.(2007)Le Boudec, McDonald, and Mundinger, Gomez-Serrano et 10, 2012] 11

  12. The Importance of Being Spatial • Mobile node state = (c, t)c = 1 … 16 (position) t ∊ R+ (age of gossip) • Time iscontinuous, I(N) = 1 • OccupancymeasureisFc(z,t) = proportion of nodesthatat location c and have age ≤ z[Age of Gossip, Chaintreau et al.(2009)] no class 16 classes Qqplots simulation vs meanfield 12

  13. What can we do with a Mean Field Interaction Model ? Large Nasymptotics, Finite Horizon fluid limit of occupancy measure (ODE) decouplingassumption (fast simulation) Issues When valid How to formulate the fluid limit Large t asymptotic Stationary approximation of occupancy measure Decoupling assumption Issues When valid 13

  14. E. L. 2. Convergence to ode 14

  15. IntensityI(N) • I(N) = expectednumber of transitions per object per time unit • A meanfieldlimitoccurswhenwere-scale time by I(N)i.e. weconsiderXN(t/I(N)) • I(N) = O(1): meanfieldlimitis in discrete time [Le Boudec et al (2007)]I(N) = O(1/N): meanfieldlimitis in continuous time [Benaïm and Le Boudec (2008)] 15

  16. The Mean Field Limit • Under verygeneral conditions (givenlater) the occupancymeasure converges, in law, to a deterministicprocess, m(t),called the meanfieldlimit • Finite State Space => ODE 16

  17. Mean Field Limit N = +∞ Stochastic system N = 1000 17

  18. Sufficient Conditions for Convergence • [Kurtz 1970], seealso [Bordenav et al 2008], [Graham 2000] • Sufficientconditonverifiable by inspection:Example: I(N) = 1/NSecond moment of number of objectsaffected in one timeslot = o(N) • Similarresultwhenmeanfieldlimitis in discrete time [Le Boudec et al 2007] 18

  19. Example: Convergence to Mean Field • Rescale time suchthat one time step = 1/N • Number of transitions per time stepisbounded by 2, thereforethereis convergence to meanfield = = = 19

  20. Formulating the Mean Field Limit • Drift = sum over all transitions ofproba of transition xDelta to system state MN(t) • Re-scale drift by intensity • Equation for meanfieldlimitis dm/dt = limit of rescaled drift • Can beautomatedhttp://icawww1.epfl.ch/IS/tsed drift = = = = 20

  21. Convergence to Mean Field E. L. E.L. • For the finite state space case, there are many simple results, oftenverifiable by inspectionFor example [Kurtz 1970] or [Benaim, Le Boudec 2008] • For the general state space, thingsmaybe more complex(fluidlimitis not an ODE, e.g. [Chaintreau et al, 2009], [Gomez-Serrano et al, 2012]) 21

  22. 3. FINITE HORIZON :FastSimulation and Decouplingassumption 22

  23. Convergence to Mean Field Limitis Equivalent to Propagation of Chaos 23

  24. Propagation of Chaos = DecouplingAssumption • (Propagation of Chaos)kobjects are asymptoticallyindependentwithcommonlawequal to the meanfieldlimit, for anyfixedk • (DecouplingAssumption) (alsocalledMean Field Approximation, or Fast Simulation)The law of one objectisasymptotically as if all otherobjectsweredrawnrandomlywith replacement fromm(t) 24

  25. The TwoInterpretations of the Mean Field Limit At any time t Thus for larget : Prob (node n is dormant) ≈ 0.3 Prob (node n is active) ≈ 0.6 Prob (node n is susceptible) ≈ 0.1 m(t) approximatesboth the occupancymeasure MN(t) the state probability for one objectat time t, drawnatrandomamongN 25

  26. « Fast Simulation » whereis the (transient) probability of a continuous time nonhomogeneous Markov process • Same ODE as meanfieldlimit, withdifferent initial condition pdf of node 2 occupancy measure (t) pdf of node 1 pdf of node 3 26

  27. The DecouplingAssumption • The evolution for one object as if the otherobjectshad a state drawnrandomly and independentlyfrom the distribution m(t) • Is validover finite horizon whenevermeanfield convergence occurs • Can beused to analyze or simulateevolution of k objects 27

  28. 4. Infinite Horizon: Fixed Point Method and Decouplingassumption 28

  29. The Fixed Point Method Decoupling assumptionsaysdistribution of prob for state of one objectiswith We are interested in stationaryregime, i.ewe do This is the « Fixed Point Method » Example: in stationaryregime: Prob (node n is dormant) ≈ 0.3 Prob (node n is active) ≈ 0.6 Prob (node n is susceptible) ≈ 0.1 Nodes m and n are independent 29

  30. Example: 802.11 Analysis, Bianchi’s Formula 802.11 single cell mi = proba one node is in backoff stage I = attempt rate  = collision proba See [Benaim and Le Boudec , 2008] for this analysis ODE for meanfieldlimit Solve for Fixed Point: Bianchi’s Fixed Point Equation [Bianchi 1998] 30

  31. 2-Step Malware, Again Same as before except for one parameter value : h = 0.1 instead of 0.3 The ODE does not converge to a unique attractor (limit cycle) The equationF(m) = 0has a unique solution (red cross) – but itisnot the stationaryregime ! 31

  32. ExampleWhereFixed Point MethodFails In stationaryregime, follows the limit cycle Assume you are in stationaryregime (simulation has run for a long time) and you observe that one node, say, is in state ‘A’ It is more likelythatis in region R Therefore, itis more likelythatsomeothernode, say, isalso in state ‘A’ This issynchronization R h=0.1 32

  33. Joint PDFs of TwoNodes in StationaryRegime Stationary point of ODE Mean of Limit of N = pdf of one node in stationary regime pdf of node 2 in stationary regime, given node 1 is A pdf of node 2 in stationary regime, given node 1 is D pdf of node 2 in stationary regime, given node 1 is S 33

  34. Whereis the Catch ? Decouplingassumptionsaysthatnodesm and n are asymptoticallyindependent There ismeanfield convergence for thisexample But wesawthatnodesmay not beasymptoticallyindependent … isthere a contradiction ? 34

  35. Markov chainisergodic • The decouplingassumptionmay not hold in stationaryregime, even for perfectlyregularmodels Mean Field Convergence ≠ mi(t) mj(t) mi(t) mj(t) 35

  36. Result 1: Fixed Point MethodHoldsunder (H) Assume that (H) ODE has a unique global stable point to which all trajectories converge Theorem [e.g. Benaim et al 2008] : The limit of stationary distribution of MNisconcentrated on thisfixed point The decouplingassumptionholds in stationaryregime 36

  37. Here:Birkhoff center = limit cycle  fixed point Theorem in [Benaim] says that the stochastic system for large N is close to the Birkhoff center, i.e. the stationary regime of ODE is a good approximation of the stationary regime of stochastic system Result 2: Birkhoff Center 37

  38. StationaryBehaviour of Mean Field Limitis not predicted by Structure of Markov Chain • MN(t)is a Markov chain on SN={(a, b, c) ≥ 0, a + b + c =1, a, b, c multiples of 1/N} • MN(t)isergodic and aperiodic • Depending on parameter, thereis or is not a limit cycle for m(t) SN(for N = 200) h = 0.3 h = 0.1

  39. Example: 802.11 withHeterogeneousNodes • [Cho et al, 2010]Two classes of nodeswithheterogeneousparameters (restransmissionprobability)Fixed point equation has a unique solution, but thisis not the stationaryprobaThere is a limit cycle 39

  40. Result 3: In the Reversible Case, the Fixed Point MethodAlways Works • DefinitionMarkov Process on enumerable state E space, with transition rates q(i,j)isreversibleiff 1. itisergodic 2. p(i) q(i,j) = p(j) q(j,i) for somep • Stationary points = fixed points • If processwithfiniteNisreversible, the stationarybehaviourisdeterminedonly by fixed points. 40

  41. A Correct Method • 1. Writedynamical system equationsin transientregime • 2. Study the stationaryregime of dynamical system • if converges to unique stationary point m*thenmakefixed point assumption • elseobjects are coupled in stationaryregime by meanfieldlimitm(t) • Hard to predictoutcome of 2 (except for reversible case) 41

  42. Conclusion • Meanfieldmodels are frequent in large scalesystems • Validity of approachisoften simple by inspection • Meanfieldisboth • ODE for fluidlimit • Fast simulation usingdecouplingassumption • Decouplingassumptionholdsatfinite horizon; may not hold in stationaryregime. • Stationaryregimeis more thanstationary points, in general (except for reversible case) 42

  43. Thank You … 43

  44. References 44

  45. 45

  46. [Gomez-Serrano et al, 2012] Gomez-Serrano J., Graham C. and Le Boudec J.-Y.The Bounded Confidence Model Of Opinion DynamicsMathematical Models and Methods in Applied Sciences, Vol. 22, Nr. 2, pp. 1150007-1--1150007-46, 2012. 2 46

  47. 47

  48. 48

More Related