Sie sind auf Seite 1von 5

Linear Programming Techniques

in R & D Project Planning*


D. H. Allen

Chemical Engineering Department,


University of Nottingham

business investment. In a vigorous R & D Table 2. Projects Ranked


In a multi-project industrial R 8 D organisation potential and existing projects
organization the selection of projects compete for the limited resourses
Project Benefit Resource
with potentially good payoffs has available. The whole range of its activities
to be fitted in with the planning of therefore needs to be coordinated and B 35 6
on-going projects as they compete planned to ensure that resources are used A 30 5
for the same expert staff laboratory C 25 3
efficiently to obtain the best possible total
equipment, pilot plant rigs and
other facilities and resources. A benefit, The use of a properly formulated
linear programming approach to and interpreted linear programming model
this planning problem takes account of the situation can help in achieving this. explore combinations of projects to find
of any flexibility possible in allocat- the set which maximizes their total
ing resources to projects, and in combined benefit within the permitted
their timing. in finding the ‘best’ total research usages. Linear programming
THE BASIC PROBLEM
plan. This is the one which maxi-
This is to select and progress projects in a is the most useful technique for this. In its
mises the potential payoff for the
whole ‘portfolio’ of projects and way which will lead to the achievement of general form it finds the values of _‘ci. . ..Y,,
makes the most efficient use of the the best total benefit, within the limits set which satisfy :
available resources. The principles by the resources available to carry them
of the formulation of such LP out. Table 1 gives numerical measures of Maximise 2 hj.X/
models and the interpretation of the imI
the individual benefits expected from N
results produced in terms of practi- subject to the constraints: Z a;jxi < c;
carrying out each of three projects A, B
cal planning are discussed. j= 1
and C, the amount of a resource (e.g.
and xi > 0 for i = 1,2 . ,117
money) which each would use, and the
total amount of this resource available. where ali, bi and Ci are constants.
INTRODUCTION A simple benefit ranking procedure could The solution of this type of problem is
N THE MANAGEMENT OF RESEARCH AND be used to select projects to the limit of the performed by well-established iterative
I development, a large R & D organi- available resource. The projects are shown procedures which are easily programmed
zation with many diverse projects presents ranked in this way in Table 2, in the order for a computer. The user is therefore
particular problems in the selection and B, A, C. Following this method, only B mainly concerned with the formulation of
timing of new projects and in planning the would be selected, using 6 units of the his problem in these terms and in the
progress of on-going ones. Besides taking resource and with a benefit of 35. A cannot subsequent interpretation of its solution.
into account the merits and needs of also be selected since it needs 5 units of
individual projects it is necessary to relate resource and only 2 are still available. The linear programming (LP) formula-
these to the allocation of the available However, it can be seen that if A and C tion of the previous trivial project selection
budget, research personnel and research are selected instead of B, the total benefit problem is given as an LP matrix in Table
facilities between the different projects. is 55 for a combined acceptable resource 3. The three projects A, B and C are
In industry, projects usually are, or should usage of 8 units. represented by ,x1, x2 and xj. These
be, selected and carried out because the variabIes are individually constrained
Methods of project selection which use between 0 and I. If xj =: 0 the corres-
company concerned hopes thereby to mathematical programming techniques
benefit in some tangible way. Expenditure ponding project is not selected: if -ui -= 1
on R & D is an investment for which a the project is selected. The objective
return is expected, as with any other Table 1. Projects for Selection. function coefficients are the project ‘bene-
fits’, and the remaining constraint governs
the allocation of the one resource. In the
optimal solution x, =y I, x2 = 0, x3 --: 1
*This article is based on a paper presented at an
Institution of Chemical Engineers Symposium on
and the maximized value of the objective
“The Management of Product and Process function is 55, indicating the selection of
Innovation”. A and C and the rejection of B.

FEBRUARY, 1974 61
Table 3. LP formulation of Project Selection Problem. a project selection model may also imply
the allocation of some other resources,
e.g. costs and equipment closely associated
Projects
with staff usage, in which case there is no
Xl x2 x3
need to include these separately.
Objective function 30 35 35 Maximum

Project Selection 1 < 1 PROJECT VERSIONS


Constraints 1 < 1 In practice it may be possible to vary the
1 =z 1
allocation of resources to, and the timing
Resource Constraint 5 6 3 < 8
of, a project if this is desirable for the
project portfolio as a whole. Such flexibility
may be built into the model by including
different ‘versions’ of a project. A project
It is of course possible for projects to be This consists of the objective function or may for example be described in alternative
‘partially’ selected by this type of model, benefit row in which the potential benefit ‘slow’, ‘normal’ and ‘fast’ versions,
i.e. xj can have intermediate values between of project Xi is the coefficient, bi, the referring to the rate at which they use
0 and 1. This does not usually occur very project selection rows with + 1 coefficients various resources. Taking extremes, a
frequently, since a characteristic of the and RHS’s, and the resource rows in slow version could include the minimum
solution of an optimization problem is for which ri, Si and fi are the amounts of the amounts of each resource needed just to
variables to take extreme values. The available resources R, S, and T which are keep a project going, and a fast version
significance of partial selection is discussed used by project xi if it is selected. The the amounts needed to complete the
in more detail later.
The measure of estimated potential
Table 4. Simple LP Project Selection Model.
project benefits must be additive in order
that the combined benefit of the total
project ‘portfolio’ may be maximized. Projects
A measure commonly used is net present Xl x2 x3 . . . x,
value (NPV), i.e. the total future benefits
discounted at an appropriate rate to a Objective Function 4 4 b, . . . 6, Maximum
single cash figure. Measures expressed as
percentage returns, e.g. DCF return, are 1 < 1
not additive and cannot be used. Selection Rows 1 < 1
1 < 1
The application of linear programming 1 < 1
in the R & D field has been discussed by
Allen and Johnson, 1,2 Bell and Read3 and 5 r2
r3 . . . rm < R
others, and represents an extension of Resource Rows Sl s2
S) . . . s, < S
previous work on the optimal allocation t1 t2
t, . . . t, < T
of funds to capital investment projects.4
The main difference between R & D and
capital investment models is that in R & D
there is usually a larger number of smaller selection rows ensure that each project, project in the shortest possible ‘crash
projects to consider which utilize several with its associated benefit and usage of time. In a multiple time-period model, to
types of resources not readily inter- resources, can only be selected once. be discussed later, there is also the flexi-
changeable, at least in the short-term. Resource rows can be concerned with bility of having alternative project versions
The degree of uncertainty in the eventual such things as financial resources (capital, starting and stopping at different times, so
outcomes of R & D projects (their benefits) revenue), facilities (pilot plant, analytical that the most appropriate timing can be
may also be higher than is usually equipment) and staff. In a financial resource selected.
acceptable with capital projects. row the coefficient of each project variable Different versions of a project will
is its appropriate cost in the time period usually have different potential benefits
DEVELOPMENT OF AN LP being considered and the RHS is the arising from the differences in timing and
PROJECT PLANNING MODEL resource usages. The resulting optimal
corresponding total budget available. In a
An LP model is of most value in situations facility constraint the coefficients are the solution will select the ‘best’ combination
where there are several types of resources proportions of available capacity (e.g. as a of project versions as well as of projects
to allocate, and where not only the percentage) used by each project. Capacity themselves, that is, the combination which
selection of new projects is considered, involves the length of time a facility is has the best total project portfolio benefit
but also alternative ways of carrying out used as well as the amount, so that using with the available resources.
new and existing projects. The future 50 per cent of a facility throughout a time In the formulation of a model a variable
planning of projects and the allocation of period is the same as far as the model is is needed for each project version. Versions
resources may also be dealt with in several concerned as using 100 per cent for half of the same project are mutually exclusive,
distinct time periods. Various features the time period. Similarly in a simple so that if one is selected the rest must be
which may be included in comprehensive staffing row the project coefficients rep- rejected. This is achieved by giving each
LP project planning models will now be resent the man-weeks (or similar measure) version of a project a $1 coefficient in
considered. they each would need in the time period that project’s selection row, as shown in
Generalizing from the earlier example, and the RHS the corresponding total man- Table 5, where xij refers to project i version
the LP matrix of Table 4 is obtained. weeks available. The allocation of staff by j. Each project version variable also has its

LONG RANGE PLANNING


62
appropriate coefficients in the objective or other resources which are partially row is required as shown in Table 7. Q is
function and the resource rows. interchangeable. R and S are the available the additional amount of resource R
available and the value of variable y gives
Table 5. Project Version Selection Rows.
the amount of it used.

Project Project Project


MULTIPLE TIME PERIOD
1 2 3
MODELS
x11 x12 x13 x21 x22 x31
Since the coefficients showing the resource
Selection Rows 1 1 1 < 1
usages of project versions are only the
1 1 < 1 averages for the time period being con-
sidered, a more detailed representation
can be obtained by replacing the planning
period by shorter distinct intervals. For
An essential requirement in formulating amounts of each which are independent, example, instead of a single planning
this type of project planning model is that I is the amount which can be used as period looking one year ahead it may be
projects are defined in such a way that either additional R or S. In a solution the better to have, say, three separate periods.
both projects and project versions are values of the interchange variables i, and The first could be the immediate three
independent of one another, so that the i, are the amounts of Z used as additional months ahead in which the detailed
selection of one project or version does R and S respectively. The R and S rows planning of projects and allocation of
not presuppose the selection of any other. state that the amount of each resource resources is considered; the next 9 months
Where actual projects do in fact interact allocated to the projects, minus the amount as the second period could be less specific,
the different alternatives should be of the interchangeable resource used, for example treating total staff as one
presented as versions of a combined cannot be more than the amount of each resource, and a third period of 12-24
project. For example, if project A can be independent resource available. The inter- months ahead could be in very general
done alone but project B can only be done change row limits the sum of the values of terms involving perhaps only financial
:f A is done as well, this is presented as a the interchange variables to the amount of resources.
combined project C having two alternative interchangeable resource available, I. The In a multiple time period model the
versions with the benefits and resource general pattern of this formulation can be number of project version variables remains
requirements of A and A+ B respectively. extended to any number of interchangeable unchanged and there is of course only one
If, as often happens in practice, resources. benefit coefficient in the ob.jective function
particularly with an on-going project,
there is no question of a project not being Table 6. Partially Interchangeable Resources.
selected, there may still be flexibility in
choosing the most appropriate version of
it to fit in with other projects. In this case Projects and Versions Interchange Variables
the ‘<’ in the selection row for the project x,, . . . . X,” ir is
is changed to ‘=‘, i.e. the project is made
mandatory so that a version of it must be R row r,, . . . rmn -.- 1 < R
selected. S row s,, . . . s,, -1 < s

R/S interchange 1 1 d I

FURTHER CONSIDERATION OF
STAFFING AND OTHER
RESOURCE CONSTRAINTS The effect of making available additional for each project version. Its resource
The most important resource in planning amounts of any resource, for example the usages in different time periods depend on
the selection and progression of projects recruitment of extra staff, may of course whether a version starts in a particular
may very well be the research staff, in be explored by changing the appropriate time period, stops in it, or continues
which case it can be advantageous to pay RHS. However, for reasons of clarity and right through it and at what rate. Thus the
more detailed attention to staffing in the interpretation it can be useful to include the resource rows in different time periods are
model. Staff with different types of facility separately in the model, particularly entirely separate, with separate project
expertise, or different levels of skills, may in multi-time period models to explore the version coefficients and RHS availabilities.
be included as separate constraints, for incentive for making available additional Similarly any other variables and rows
example chemists and chemical engineers, amounts of resources in the future. For connected with resources (e.g. interchange)
or technical officers, experimental officers each resource an additional variable and are presented independently in each period.
and laboratory assistants. If in practice
some staff are able to function in more
Table. 7 Additional Resource (Recruitment).
than one of the specified categories, this
can be reproduced in the model by means
of staff interchange rows and variables. Projects and Versions Additional R Variable
The representation can also be extended to x,, . . . X,” Q
other types of partially interchangeable
resources. R row r,, . . . . rmn .~ 1 <R
Table 6 shows the representation of the Additional R row 1 <c-I
situation wherethereare two staff categories

FEBRUARY, 1974 63
A COMPLETE PROJECT a similar budget row but staff’ are not BASIC CONSIDERATIONS IN
PLANNING MODEL broken down into categories this far FORMULATING AN LP PLANNING
A project planning model should reproduce ahead, so that a total staff row is adequate. MODEL
as realistically as possible the salient The solution of the LP matrix, with the These are of three types depending on the
features of the actual R & D situation and aid of a computer program of a type level in the company at which they should
there are no rigid rules for its formulation, readily available with most computers, originate. The first consideration is that
except perhaps that it should be able to selects the set of values for the variables the model should reflect the corporate
select not only projects but also the most within the ranges permitted by the con- objectives of the company. It is only
appropriate plan (version) for a project. straints which maximise the value of the possible to optimize on one objective
To illustrate the form of a complete model, objective function. Tt thus selects the function; if there are multiple objectives,
some of the features so far discussed are projects and versions (the portfolio) from one major objective must be chosen for
put together in the LP project planning those offerred which achieves the highest maximization as the objective function
matrix of Table 8. portfolio potential benefit and indicates and the other objectives maintained at

Table 8. A Complete LP Project Planning Model.

TP-1 Staff
Project 1 Project 2 Project m Interchange
x11 . . . x10 '21 . . . ‘lb X ml . . ‘177” ir i2

Objective Function Project version benefit coefficients = Max

Selection Project 1 1 . . . . 1 < 1


Rows Project 2 1 . . ..I 1 1
. . .
Project m 1 . . . . . 1 < 1

Budget Budget usage coefficients < Cl


Time Staff 1 Staff 1 usage coefficients -1 < Sll

Period 1 Staff 2 Staff 2 usage coefficients -1 < s12

Staff l/2 1 1 < ‘I

Time Budget Budget usage coefficients < c2

Period 2 Total Staff Total staff usage coefficients < s2

It contains data on m projects and the corresponding allocation of resources acceptable levels by expressing them as
. - .
alternative versions are offered for each. between projects. The optimal solution constraints. ror example, total NPV at
The objective function has coefficients for also provides marginal economic informa- I5 per cent could be maximized subject to
each project version representing the tion (shadow prices) on the effect of a minimum constraint on the value of
potential benefit of that version of the changing the amount of any resource sales turnover at a specific time in the
project. There is a project selection row available and of selecting alternative future. This would guard against any
for each project, consisting of $1 co- projects or versions. tendency using NPV to concentrate on
efficients for each version of that project Since an LP planning model is intended short-term projects at the expense of
and a RHS also of + 1. Project 2 is to be a tool to assist research management, longer-term ones, and still provide the
mandatory, so that a version of it must be the input and output of the model should most satisfactory way of meeting the
selected; the other projects are not be in forms familiar to the management combined objectives. In this case the
mandatory, so that they can be rejected or rather than to a mathematician. To estimated contribution of each project
a version selected, depending on what is achieve this, additional computer programs version to turnover at the future date
best for the project portfolio as a whole. may be written around the LP program to would become the coefficients of this extra
Two time periods are considered from the generate the LP matrix from the data constraint row.
point of view of the allocation of resources. provided and to interpret the LP solution The next considerations are at the level
Tn the immediate one the resources into a detailed plan, as illustrated in of the R & D management and are
considered are a budget and two categories Table 9. concerned with identifying the resources
of research staff. The project version
coefficients of the two staff rows are the
Table 9. Requirements of LP Project Planning Software.
requirements of each version if selected,
and the RHS’s are the totals available.
There are some staff who can be effective 1. LP matrix generation from raw data.
in either category and these are included 2. Printout of original data in meaningful form as a check for the users.
separately as the RHS of a third staff row. 3. Identification list of matrix columns and rows for users’ benefit.
The staff interchange variables which 4. Full printout of generated LP matrix.
5. Details of LP iterations.
indicate the way these staff are used each 6. Listing of full LP optimal solution.
have coefficients of - 1 in the corresponding 7. Interpretation of solution for users in terms of individual projects, resource usages, etc.
staff category row and +l in the third 8. Interpretation of shadow prices for users in terms of ‘next best’ alternatives, etc.
staff row. The second time period contains

64 LONG RANGE PLANNING


whose inclusion in the model are necessary, An alternative approach altogether is to management techniques, their use should
the most appropriate time periods to use solve the model by integer programming be as an aid to the normal management
and then the actual data on the amounts of which forces the variables in the solution decision processes, and not a substitute for
resources available in these time periods. to have integer values, i.e. project version them. An LP model draws specific and
Finally, the actual projects themselves variables to be 0 or 1. This is more logical conclusions from the input data
are considered at the level of individual complicated computationally and, al- provided, but it must be borne in mind
project supervision. This is to resolve though it produces a ‘tidier’ solution, this that this data is in any case largely
questions of the project versions to be may be unnecessarily distorted by not subjective.
included and the estimation of benefit and providing information on the ‘marginal’ In order to be of any value the formula-
resource usage data for each project project or on the omission of desirable yet tion of an LP model must be acceptable to
version. The estimation of benefit data is feasible versions from the input. the research management and research
usually the most difficult and time- staff as being a realistic representation of
consuming part of using an LP project UNCERTAINTY IN DATA their organisation’s situation and activities.
planning model. However, this type of Uncertainty is of course a major factor in The results obtained must likewise be
information is needed for any optimizing R & D and cannot be ignored. In a project explainable and acceptable in terms of
approach to planning and is not a specific planning model there may be uncertainty real projects and real people. For example,
difficulty of LP models. In arriving at the in the availability of resources (the RHS’s) although staff may be treated as an
potential benefit of a project version it is its the potential benefits of project versions impersonal resource in the model, the
incremental effect which should be con- (objective function coefficients) and in the results need to be interpreted in terms of
sidered, that is, the difference between the resource requirements of project versions actual individuals, taking into account
future situation if the project is selected (matrix body coefficients). Since it is a wherever possible their individual pref-
and completed compared with situation if it simple matter to rerun a model with erences.
is not. Where the same results could be changes in data, particularly if parametric Although the initial effort required to
achieved by alternative means, for example, programming facilities are available, a introduce an LP project planning system
by licensing, it is the benefit relative to this comprehensive sensitivity analysis can be may be considerable, particularly in
next best alternative that is required. carried out to provide a ‘feel’ for how the gathering the data, its routine use involves
‘best’ portfolio responds to variations in a much lower level of effort. Main ‘up-
THE PROBLEM OF PARTIALLY any input data, i.e. its ‘robustness’ in the dating’ runs need to be carried out at
SELECTEDPROJECTS face of uncertainty. intervals not exceeding the length of the
Using a straightforward linear program- The largest uncertainty is usually in initial time period in the model. For these,
ming method of solution, the ‘optimal’ estimates of project version potential entirely new data will only be required for
project portfolio produced by a model may benefits. It may be more realistic to a few new projects. Data for the rest will
contain partially selected project versions, present these in the form of possible only need revising to take account of the
i.e. the value of one or more version ranges, probability distributions or discrete effect of progress and changes since the
variables may lie between 0 and 1 instead alternative values with associated prob- last main run. Also for ongoing projects
of being 0 or 1. There are a number of abilities. These may be reduced to single- the number of versions in the model can
practical reasons why this occurs. A single valued ‘expectations’ which then become usually be reduced in successive updating
partially selected project version may be the objective function coefficients. An runs, lessening the amount of data
the last ‘marginal’ one which uses up a alternative method also takes into account revision needed. Runs for sensitivity
residual resource to its limit. Thus the the variances (as a measure of the spread analysis and similar exploration based on
project obviously should not be selected of the distributions) as well as the expected the current model do not require extensive
unless sufficient resources can be made values in arriving at the objective function changes to data.
available to do it properly, since a partial coefficients. Since an LP model needs to be tailored
project is meaningless. Another approach, which avoids the to the particular needs of the organisation,
It may also seem obvious to say that the somewhat arbitrary reduction of a potential it is advisable to start with a relatively
optimal solution can only consist of benefit distribution to a single figure, is to simple model to gain experience, and to
project versions which have been included use a method based on Monte Carlo add to it as and when the need is recognized.
in the model. If the most appropriate simulation. In this a large number of This approach, rather than jumping in
versions of projects are not offerred for LP runs are done, in each of which values straightaway with a more complex model
selection they cannot be selected, and this for the benefit coefficients are selected needing more data, helps to ensure that
can also be the cause of partially selected from their distributions in such a way the size of the model and its data require-
project versions. If for example the ‘normal’ that their frequencies of selection cor- ments are kept to the minimum commen-
version of a project is, say, 0.6 selected and respond to their probability distributions. surate with its usefulness. ??
the ‘fast’ version 0.4 selected, this means A set of ‘optimum’ project portfolios is
that the best version of the project may be obtained corresponding to different
one in which the rate of resource usage possible benefit ‘futures’. The one finally REFERENCES
(and benefit) is somewhere between the decided upon should where possible be (1) D. H. Allen and T. F. N. Johnson, The
‘normal’ and ‘fast’ versions offerred. robust in the face of benefit uncertainty, Chemical Engineer, Volume 241 CE278
Rerunning with a suitable intermediate (September 1970).
i.e. its frequency of appearance in the
version included would confirm this. simulation solutions should be high. (2) D. H. Allen and T. F. N. Johnson, R B D
Occasionally a partially selected project Management, Volume 1, Number 2, p. 95

version may not apparently fit either of USEFULNESS OF LP PROJECT (1971).

these explanations. A useful procedure PLANNING MODELS (3) D. C. Bell and A. W. Read R 8 D Manage-
ment, Volume 1, Number 1, p. 35 (1970).
here is to rerun omitting the version which LP project planning models have, to the
(4) H. M. Weingartner, Mathematical Pro-
was partially selected, as the resulting author’s knowledge, recently been tried in gramming and the Analysis of Capital
reallocation and reselection may result in several R & D organizations with varying Budgeting Problems. Markham, Chicago
a practical optimal solution. degrees of success. Like other quantitative (1967).

FEBRUARY, 1974 65

Das könnte Ihnen auch gefallen