Iterative Transient State Distribution Calculation in Semi-Markov Processes Nick Dingle njd200@doc.ic.ac.uk Department of Computing, Imperial College London, South Kensington Campus, London SW7 2AZ. PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.1 Motivation Have focused mainly on passage time density and quantile calculation in previous work Other measures of interest, e.g. transient state distributions System may never reach steady-state due to resets etc, May be interested in knowing the probability of being in a certain state at a given moment in time PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.2 Passage Times PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.3 Transient State Distributions 0.04 Transient solution Steady-state solution 0.035 0.03 Probability 0.025 0.02 0.015 0.01 0.005 0 0 20 40 60 80 100 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.4 Motivation Existing transformation of passage times → transient distributions computational expensive To overcome this, we seek to apply our iterative passage time algorithm directly to transient state distribution calculation Work presented here is the contribution of the SI of our PMEO-PDS 2003 paper recently submitted to FGCS PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.5 Semi-Markov Processes A more expressive generalisation of Markov processes where state sojourn times can be generally distributed PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.6 Semi-Markov Processes A more expressive generalisation of Markov processes where state sojourn times can be generally distributed An N -state SMP is characterised by: PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.6 Semi-Markov Processes A more expressive generalisation of Markov processes where state sojourn times can be generally distributed An N -state SMP is characterised by: an N × N matrix P where pij is the probability of moving from state i to state j , PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.6 Semi-Markov Processes A more expressive generalisation of Markov processes where state sojourn times can be generally distributed An N -state SMP is characterised by: an N × N matrix P where pij is the probability of moving from state i to state j , an N × N matrix H where hij (t) is the distribution function of the sojourn time of the process in state i, given that its going to state j PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.6 Semi-Markov Processes A more expressive generalisation of Markov processes where state sojourn times can be generally distributed An N -state SMP is characterised by: an N × N matrix P where pij is the probability of moving from state i to state j , an N × N matrix H where hij (t) is the distribution function of the sojourn time of the process in state i, given that its going to state j We def ine the kernel R(i, j, t) of an SMP as: R(i, j, t) = pij Hij (t) PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.6 Transient State Distributions I The transient state distribution πij (t) is: πij (t) = IP(Z(t) = j | Z(0) = i) where Z(t) is the state of the SMP at time t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.7 Transient State Distributions I The transient state distribution πij (t) is: πij (t) = IP(Z(t) = j | Z(0) = i) where Z(t) is the state of the SMP at time t There is a translation from passage time densities to transient state distributions in LT form (Pyke): ∗ (s) 1 − h 1 ∗ i πij (s) = if i = j s 1 − Lii (s) ∗ ∗ πij (s) = Lij (s)πjj (s) if i 6= j ∗ (s) is the Laplace transform of π (t) and where πij ij P ∗ (s) h∗i (s) = k rik PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.7 Transient State Distributions II For multiple target states: πi∗~j (s) = X ∗ πik (s) k∈~j PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.8 Transient State Distributions II For multiple target states: πi∗~j (s) = X ∗ πik (s) k∈~j To construct πi∗~j (s) using the translation is computationally expensive: for a vector of target states ~j , need 2|~j| − 1 passage time quantities Lik (s), requires the solution of |~j| linear systems PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.8 Transient State Distributions II For multiple target states: πi∗~j (s) = X ∗ πik (s) k∈~j To construct πi∗~j (s) using the translation is computationally expensive: for a vector of target states ~j , need 2|~j| − 1 passage time quantities Lik (s), requires the solution of |~j| linear systems This motivates our development of a new transient state distribution formula for multiple target states PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.8 Iterative Transient Algorithm I From Pyke we derive the following expression for πij (t): πij (t) = δij F i (t) + N Z X k=1 0 t R(i, k, t − τ ) πkj (τ ) dτ where: δij = 1 if i = j and 0 otherwise F i (t) is the reliability function of the sojourn time distribution in state i PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.9 Iterative Transient Algorithm I From Pyke we derive the following expression for πij (t): πij (t) = δij F i (t) + N Z X k=1 0 t R(i, k, t − τ ) πkj (τ ) dτ where: δij = 1 if i = j and 0 otherwise F i (t) is the reliability function of the sojourn time distribution in state i PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.9 Iterative Transient Algorithm I From Pyke we derive the following expression for πij (t): πij (t) = δij F i (t) + N Z X k=1 0 t R(i, k, t − τ ) πkj (τ ) dτ where: δij = 1 if i = j and 0 otherwise F i (t) is the reliability function of the sojourn time distribution in state i PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.9 Iterative Transient Algorithm II Transforming the convolution into the Laplace domain and generalising to multiple target states ~j : πi∗~j (s) = ∗ δi∈~j F i (s) + N X ∗ rik (s)πk∗~j (s) k=1 where δi∈~j = 1 if i ∈ ~j and 0 otherwise PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.10 Iterative Transient Algorithm IV E.g. when ~j = {1, 3}: ∗ (s) 1 − r11 ∗ (s) −r12 ··· ∗ (s) −r1N ∗ (s) −r21 ∗ (s) 1 − r22 ··· ∗ (s) −r2N ∗ (s) −r31 ∗ (s) −r32 ··· ∗ (s) −r3N .. . .. . .. .. . ∗ (s) −rN 2 ∗ (s) −rN 2 ··· PASTA 2004 Edinburgh . ∗ 1 − rN N (s) njd200@doc.ic.ac.uk π ∗~ (s) 1j π ∗~ (s) 2j π ∗~ (s) 3j . .. π ∗ ~ (s) Nj = June 2004 ∗ F 1 (s) 0 ∗ F 3 (s) .. . 0 – p.11 Iterative Transient Algorithm V In general: (r) (I − U) π~ (s) j =v where: ∗ (s) matrix U has elements upq = rpq (r) vector π~ (s) has elements j ³ ´ π ∗~ (s), π ∗~ (s), . . . , π ∗ ~ (s) 1j 2j Nj vector v has elements vi = PASTA 2004 Edinburgh njd200@doc.ic.ac.uk ∗ δi∈~j F i (s) June 2004 – p.12 Iterative Transient Algorithm VI Therefore: (r) π~ (s) = (I − U)−1 v j ´ ³ = I + U + U 2 + U3 · · · v PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.13 Iterative Transient Algorithm VI Therefore: (r) π~ (s) = (I − U)−1 v j ´ ³ = I + U + U 2 + U3 · · · v Approximate this as: ´ ³ (r) π~ (s) = I + U + U2 + · · · + Ur v j PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.13 Iterative Transient Algorithm VI Therefore: (r) π~ (s) = (I − U)−1 v j ´ ³ = I + U + U 2 + U3 · · · v Approximate this as: ´ ³ (r) π~ (s) = I + U + U2 + · · · + Ur v j Can generalise to multiple start states by employing an initial vector α: ´ ³ (r) π~~ (s) = α I + U + U2 + · · · + Ur v ij PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.13 Example where X ∼ exp(2.0) and Y ∼ det(2.0) PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.14 Example 0.8 Analytic solution 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Example 0.8 Analytic solution 1 iteration 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Example 0.8 Analytic solution 1 iteration 2 iterations 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Example 0.8 Analytic solution 1 iteration 2 iterations 4 iterations 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Example 0.8 Analytic solution 1 iteration 2 iterations 4 iterations 6 iterations 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Example 0.8 Analytic solution 1 iteration 2 iterations 4 iterations 6 iterations 8 iterations 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Example 0.8 Analytic solution 1 iteration 2 iterations 4 iterations 6 iterations 8 iterations 10 iterations 0.7 0.6 Probability 0.5 0.4 0.3 0.2 0.1 0 0 2 4 6 8 10 12 Time, t PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.15 Conclusion Presented an iterative transient state distribution calculation method which is more eff icient than previous passage time-based transformations Builds on our existing iterative passage time algorithm Similarities with Markovian uniformization Applicable to both Markov and semi-Markov processes Applied to large SMPs generated from SM-SPNs (PMEO) and SPAs (PASTA) PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.16 PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.17 First Passage Times First passage time from state i to target states ~j in an SMP: Pi~j = inf{u > 0 : Z(u) ∈ ~j, N (u) > 0, Z(0) = i} Z(t) is the state of the SMP at time t N (u) is the number of state transitions by time u Laplace transform of Pi~j density can be computed as: Li~j (s) = X ∗ rik (s)Lk~j (s) + k∈ /~j X ∗ rik (s) k∈~j ∗ (s) is the LST of R(i, k, t) where rik PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.18 Iterative PT Algorithm I Generates successively better approximations to the passage time quantity Pi~j (r) ij P ~ is the conditional passage time for the system to reach any state in ~j in r transitions: (r) P ~ = inf{u > 0 : Z(u) ∈ ~j, 0 < N (u) ≤ r, Z(0) = i} ij (r) P~ ij → Pi~j as r → ∞ PASTA 2004 Edinburgh njd200@doc.ic.ac.uk June 2004 – p.19 Iterative PT Algorithm II (r) P~ : ij Calculate Laplace transform of rth iteration, r X (r) L ~ (s) = contribution from mth transition matrix ij = m=1 r X UU 0 m−1 e~j m=1 ∗ (s) and U0 has where matrix U has elements upq = rpq states in ~j made absorbing Can generalise to multiple source states ~i: r X (r) 0 m−1 L~~ (s) = αUU e~j ij PASTA 2004 Edinburgh m=1 njd200@doc.ic.ac.uk June 2004 – p.20 F i(t) The LT of F i (t), ∗ F i (s), is generated from h∗i (s): ∗ F i (s) PASTA 2004 Edinburgh 1 − h∗i (s) = s njd200@doc.ic.ac.uk June 2004 – p.21