Chapter 3
The Bayesian Approach
Bayesian inference methods [9] provide a well-studied toolkit for calculating
a distribution of a quantity of interest given observed evidence (measurements).
As such, they are well-suited for calculating a probability distribution of the final
location of the aircraft given the data available from the Inmarsat satellite
communication system. The resulting probability distribution is essential to
prioritise search efforts. In this chapter, we provide a brief introduction to Bayesian
methods. We assume a reasonable background in probability theory; the interested
reader is referred to excellent resources such as [8, 9, 10, 18, 36, 39, 43] if further
detail is desired.
The required probability density function (pdf) is the probability of the aircraft
location given the available data. Bayes’ rule defines a method to calculate this pdf
using prior information, including knowledge of how aircraft move, and a model of
how the measured data relate to the aircraft location and velocity. Mathematically,
Bayes’ rule is
p(x, z)
p(x|z) = (3.1)
p(z)
p(z|x) p(x)
= (3.2)
p(z)
p(z|x) p(x)
= (3.3)
p(z|x ) p(x )dx
where the elements are:
1. x is the random variable, or the state, which is the quantity of interest (e.g., the
position of the aircraft);
2. z is the measurement (e.g., the Inmarsat satellite communication data, which
provides some form of positional data);
3. p(x) is the prior pdf of the state (not incorporating the measurement, e.g., based
on historical data);
© Commonwealth of Australia 2016 11
S. Davey et al., Bayesian Methods in the Search for MH370, SpringerBriefs
in Electrical and Computer Engineering, DOI 10.1007/978-981-10-0379-0_3
12 3 The Bayesian Approach
4. p(z|x) is the pdf of the measurement conditioned on the state (e.g., this may be
constructed by observing the distribution of measurements in cases where the
state is known);
5. p(x|z) is the conditional pdf of interest (the posterior pdf), describing the distribu-
tion of state (e.g., aircraft location) taking into account the observed measurement.
The posterior probability density is based on the accumulated Inmarsat satellite
communications data as well as all available contextual knowledge on the sensor
characteristics, aircraft dynamic behaviour and environmental conditions and con-
straints. The method is based on the state space approach to time series modelling.
Here, attention is focused on the state vector of a system. The state vector contains
all relevant information required to describe the system under investigation at a given
point in time. For example, in radar tracking problems this information would typi-
cally be related to the kinematic characteristics of the aircraft, such as position,
altitude, speed, and heading. The measurement vector represents noisy observations
that are related to the state vector. For example, the distance and bearing angle
between the sensor and the object being measured. The state-space approach is
convenient for handling multivariate data and nonlinear, non-Gaussian processes;
it provides a significant advantage over traditional time series techniques for these
problems; and has been extensively used in many diverse applications over the last
50 years [7]. An excellent summary of Bayesian techniques for state space models
is given by [36].
In order to proceed, two models are required: first, the measurement model relates
the noisy measurements to the state; and second, the system or dynamic model
describes the evolution of the state with time. The measurement model used for BTO
and BFO metadata is defined in a probabilistic form in Chap. 5. The dynamic model
used to define the behaviour of the aircraft is defined in Chaps. 6 and 7.
If the measurement model and the system model are both linear and Gaussian,
the optimal estimate can be calculated in closed form using the Kalman filter [25].
If either the system or measurement model is nonlinear or non-Gaussian, the pos-
terior pdf will be non-Gaussian and standard analysis with a Kalman filter will
be suboptimal. This results in the need for approximate computational strategies
and the approach adopted in this study is introduced in this chapter. The appli-
cation of the measurement and dynamics models to this approach is described in
Chap. 8. The computational approach proceeds in essentially two stages: prediction
and update. The prediction stage uses the aircraft dynamic model to step from the
state pdf at one time to the pdf at the next time. The state is subject to unknown
disturbances, modeled as random noise, and also unknown control inputs, such as
turn commands, and so prediction generally translates, deforms, and broadens the
state pdf. The update operation uses the latest measurement to modify (typically to
tighten) the prediction pdf. This is achieved using Bayes theorem, (3.3), which is the
mechanism for updating knowledge about the state in the light of extra information
from new data.
3.1 The Problem and its Conceptual Solution 13
3.1 The Problem and its Conceptual Solution
To define the problem of nonlinear filtering, let us introduce the state vector
x(t) ∈ Rn , where n is the dimension of the state vector. Here t is continuous-valued
time. The state evolution is best described using a continuous-time stochastic differ-
ential equation, sometimes specifically referred to as an Itô differential equation [23].
However, it is often more convenient to sample this at discrete time instants, in which
case xk ≡ x (tk ) represents the state at the kth discrete sample time. The elapsed time
between samples k = tk − tk−1 is not necessarily constant. The state is assumed to
evolve according to a continuous-time stochastic model:
dx(t) = f (x(t), dv(t), t, dt) , (3.4)
where f(·) is a known, possibly nonlinear deterministic function of the state and v(t)
is referred to as a process noise sequence, which caters for random disturbances in
the aircraft motion.
A sensor collects measurements, which are a possibly nonlinear function of the
state. Measurements occur at times tk , for k = {1, 2, . . . K }. The kth measurement
is denoted zk ∈ Rm where m is the dimension of the measurement vector. The mea-
surements are related to the state via the measurement equation:
zk = hk (xk , wk ) , (3.5)
where hk (·) is a known, possibly nonlinear function and wk is a measurement noise
sequence. The noise sequences v(t) and wk will be assumed to be white, with known
probability density functions and mutually independent. The initial state is assumed
to have a known pdf p (x0 ) and also to be independent of noise sequences.
We seek estimates of xk based on the sequence of all available measurements up to
time tk , defining the measurement history Zk {z1 , . . . zk }. From a Bayesian perspec-
tive, the problem is to recursively construct the posterior pdf p (xk |Zk ). In principle,
the pdf p (xk |Zk ) may be obtained recursively in two stages: prediction and update.
The prediction stage steps from the pdf of x at time tk−1 , p (xk−1 |Zk−1 ), to the pdf
at the next time, p (xk |Zk−1 ), not incorporating any new measurements. The update
stage takes the predicted pdf p (xk |Zk−1 ) and incorporates the new measurement zk
occurring at time tk to obtain the updated pdf p (xk |Zk ). If there is a requirement
to evaluate the pdf at time t for which there is no measurement then this pdf is the
predicted pdf and no update step needs to be performed.
3.1.1 Prediction
The prediction stage involves using the system model (3.4) to obtain the prediction
density of the state at time step k via the Chapman–Kolmogorov equation:
14 3 The Bayesian Approach
p (xk |Zk−1 ) = p (xk |xk−1 , Zk−1 ) p (xk−1 |Zk−1 ) dxk−1 ,
= p (xk |xk−1 ) p (xk−1 |Zk−1 ) dxk−1 . (3.6)
The first line of (3.6) is a statement of the law of total probability. The sim-
plification p (xk |xk−1 , Zk−1 ) = p (xk |xk−1 ) used to progress from the first line of
(3.6) to the second applies because (3.4) describes a Markov process of order one.
The probabilistic model of the state evolution, p (xk |xk−1 ), is defined by the system
equation (3.4) and the known statistics of v(t).
3.1.2 Update
At time tk a measurement zk becomes available and the update stage is carried out.
This involves an update of the prediction (or prior) pdf via Bayes’ rule:
p (xk |Zk ) = p (xk |zk , Zk−1 )
p (zk |xk , Zk−1 ) p (xk |Zk−1 )
=
p (zk |Zk−1 )
p (zk |xk ) p (xk |Zk−1 )
= , (3.7)
p (zk |Zk−1 )
where conditional independence has been used to write the likelihood function
p (zk |xk , Zk−1 ) = p (zk |xk ), which is defined by the measurement model (3.5) and
the known statistics of wk . The normalizing constant on the denominator can be
expanded as
p (zk |Zk−1 ) = p (zk |xk ) p (xk |Zk−1 ) dxk . (3.8)
In the update stage (3.7), the measurement zk is used to modify the prior density to
obtain the required posterior density of the current state.
Note that there is no requirement for all of the measurements to have the same
statistical model or even contain the same type of information. For example, there
could be multiple sensors operating on different modalities. For simplicity, we have
not introduced explicit notation to change the measurement pdf for each k. For the
accident flight three different types of measurement have been used. As discussed in
Chap. 5, the satellite communications messages consist of R-channel and C-channel
messages that have differing information content. Another quite different form of
measurement is the areas of the ocean floor that have been searched without locating
the aircraft and the debris that has been recovered. This measurement and its potential
use to refine the ongoing search are discussed in Chap. 11.
3.1 The Problem and its Conceptual Solution 15
The recurrence relations (3.6) and (3.7) form the basis for the optimal Bayesian
solution. The recursive propagation of the posterior density, given by (3.6) and (3.7),
is only a conceptual solution in the sense that in general it cannot be determined
analytically. In most practical situations the analytic solution of (3.7) and (3.8) is
intractable and numerical approximations have to be used. This has been a topic of
significant research effort over the past 20 years [1, 20, 33]; a general overview of
the method is presented next.
3.2 The Particle Filter
In the linear Gaussian case, the pdfs for p (v(t)), p (wk ), p (x0 ) are all Gaussian and
the functions f(·) and h(·) are linear. It can then be easily shown that the posterior
p (xk |Zk ) is also Gaussian and all of these pdfs can be summarised completely by
their means and covariances. The Kalman filter is an algorithm that defines recursions
for the mean and covariance of p (xk |Zk ) in terms of the means and covariances of
the prior and noise processes. However, in general, the posterior does not take the
same functional form as the prior and indeed it is not possible to even write a closed
form expression for p (xk |Zk ). In this case an approximate solution is required.
The solution used for the MH370 search definition is referred to as the particle filter
and is a numerical approximation based on random sampling.
The fundamental concept in the particle filter is to approximate the pdf p (xk |Zk )
as a weighted combination of sample points
P
p p
p (xk |Zk ) ≈ wk δ xk − xk , (3.9)
p=1
p p
where the wk are referred to as weights and sum to unity, and the xk are referred
to as particles. The convergence properties of this approximation in the limit as the
number of particles P increases have been well studied, for example [14, 21]. Given
this approximate pdf, it is simple to evaluate the expectation of any nonlinear function
of the state, such as
P
p p
E g (xk ) |Zk ≡ g (xk ) p (xk |Zk ) dxk ≈ wk g xk . (3.10)
p=1
The approximation of an integral using sample points as above is referred to as
Monte Carlo integration and can be applied to both the Chapman–Kolmogorov pre-
diction (3.6) and the Bayesian update (3.7).
The particle filter is an algorithm that provides a mechanism to recursively create
a set of weighted particles approximating p (xk |Zk ) starting from a previous set of
weighted particles approximating p (xk−1 |Zk−1 ). It does this in two stages: first it
16 3 The Bayesian Approach
p p
moves the particle sample points xk−1 → xk to new locations using a pdf referred to
as a proposal distribution, which is a tractable approximation of the pdf of interest.
Second, it determines new particle weights to correct for the difference between the
proposal and the true pdf. This process is known as importance sampling [1, 33].
The proposal distribution is a critical component of the particle filter. It is a
function chosen by the designer subject to relatively loose constraints. Importantly,
the proposal distribution must cover all of the state space where the true distribution
is non-zero and its tails should be heavier than the tails of the true distribution. If the
p
proposal is chosen poorly then many of the particles xk will be assigned very low
weights and the filter efficiency will be low: a large number of particles will be
required for satisfactory performance. A common version of the particle filter is the
Sample-Importance-Resample (SIR) particle filter that uses the system dynamics as a
proposal distribution. The SIR is popular because it is often relatively straightforward
to sample from the dynamics and because the weight update equation is very simple
when the dynamics is used as the proposal. The filter used in this book is a form of
SIR particle filter.
p p
For the SIR particle filter, for each particle xk−1 a new xk is drawn from the tran-
p
sition density p(xk |xk−1 ), and weights are updated by scaling the previous weights
by the current measurement likelihood and re-normalising,
p p
wk = (Wk )−1 p zk |xk wk−1 ,
p
(3.11)
p p
where the normalising term is Wk = Pp=1 p zk |xk wk−1 .
A key difficulty in particle filters is the issue of degeneracy, i.e., over time, many
weights tend toward zero, and the corresponding particles are of little use. Resam-
pling is used to combat this difficulty. The simplest approach is draw P new particles
from the approximate distribution (3.9), such that particles with very large weights
are likely to be replicated many times over, and those with very small weights
are unlikely to be sampled. A variety of methods are possible, and can be found
in [1, 33]. The sampling method used in this study is detailed in Chap. 8.
3.3 Rao–Blackwellised Particle Filter
One of the challenges in implementing a particle filter is that the number of particles
required to make a good approximation to the desired posterior pdf can grow expo-
nentially with the dimension of the state space. In some circumstances, it is possible
to mitigate this by incorporating an analytic representation of the distribution of part
of the state given a sample of the remainder of the state. For example, suppose that
the measurement function can be decomposed into two parts
zk = hk1 xk1 + hk2 xk2 + wk , (3.12)
3.3 Rao–Blackwellised Particle Filter 17
where xk1 and xk2 are disjoint sub-vectors of the state xk . In this case we can write
p xk1 , xk2 |zt , Zt−1 = p xk1 |zt , Zt−1 p xk2 |xk1 , zt , Zt−1 . (3.13)
The two
densities above can be estimated using different filters. When the function
hk2 xk2 is linear and the noise is Gaussian, the second density p xk2|xk1 , zt , Zt−1
can be estimated using a Kalman filter, even if the first function hk1 xk1 is nonlinear.
The state vector that needs to be sampled is then xk1 not [xk1 , xk2 ] and the sampling
process can use fewer samples for a given degree of accuracy.
When a particle filter is used for the nonlinear
part of the measurement
problem,
the conditioning of the second state density p xk2 |xk1 , zt , Zt−1 leads to a separate
Kalman filter for each particle. Each Kalman filter uses the sampled value of
the sub-state xk1 as though it were the truth. This arrangement is referred to as a
Rao–Blackwellised particle filter [15, 29, 38].
Open Access This chapter is distributed under the terms of the Creative Commons
Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/
by-nc/4.0/), which permits any noncommercial use, duplication, adaptation, distribution and
reproduction in any medium or format, as long as you give appropriate credit to the original
author(s) and the source, a link is provided to the Creative Commons license and any changes
made are indicated.
The images or other third party material in this chapter are included in the work’s Creative
Commons license, unless indicated otherwise in the credit line; if such material is not included
in the work’s Creative Commons license and the respective action is not permitted by statutory
regulation, users will need to obtain permission from the license holder to duplicate, adapt or
reproduce the material.
