Demonstrator Technical Description:
Decision Support Functions

The ultimate objective of a computer based decision support system for urban traffic and air quality management is to improve planning and operational decision making processes by providing useful and scientifically sound information to the actors involved in these processes, including public officials, planners and scientists, and possibly the general public. This information must be:

• timely in relation to the dynamics of decision problem
• accurate in relation to the information requirements
• directly understandable and usable
• easily obtainable, i.e., cheap in relation to the problems implied costs.

Decision support is a very broad concept, and involves both rather descriptive information systems, that just demonstrate alternatives, as well as more formal normative, prescriptive optimization approaches that design them. Any decision problem can be understood as revolving around a choice between alternatives. These alternatives are analyzed and ultimately ranked according to a number of criteria by which they can be compared; these criteria are checked against the objectives and constraints (our expectations), involving possible trade-offs between conflicting objectives. An alternative that meets the constraints and scores highest on the objectives is then chosen. If no such alternative exists in the choice set, the constraints have to be relaxed, criteria have to be deleted (or possibly added), and the trade-offs redefined.

However, the key to an optimal choice is in having a set of options to choose from that does indeed contain an optimal solution. Thus, the generation or design of alternatives is a most important, if not the most important step. In a modeling framework, this means that the generation of scenarios must be easy so that a sufficient repertoire of choices can be drawn upon.

The selection process is then based on a comparative analysis of the ranking and elimination of (infeasible) alternatives from this set. For spatially distributed and usually dynamic models -- natural resource management problems most commonly fall into this category -- this process is further complicated, since the number of dimensions (or criteria) that can be used to describe each alternative is potentially very large. Since only a relatively small number of criteria can usefully be compared at any one time (due to the limits of the human brain rather than computers), it seems important to be able to choose almost any subset of criteria out of this potentially very large set of criteria for further analysis, and modify this selection if required.

DSS approach

In SIMTRAP, the decision support approach chosen is primarily constrained by the characteristics of the underlying system. These are:

• dynamic, with a typical time resolution in the order of minutes;
• spatially distributed, with spatial resolution ranging from the street level (meters) to the regional air quality grid (km);
• highly non-linear and involving time-delays and memory in the cause-effect relationships.

These problem characteristics preclude any straight forward optimization approach.

Consequently, SIMTRAP uses an approach centered on

• Scenario Analysis
• comparative evaluation of scenarios
• discrete multi-criteria optimization.

Scenario Analysis

In a DSS framework, Scenario Analysis supports the user to explore a number of WHAT -- IF questions. The scenario is the set of initial conditions and driving variables (including any explicit decision variables) that completely characterize the system behavior, which is expressed as a set of output or performance variables.

Decision Variables

The decision parameters the user can set to define a scenario include:

• meteorological conditions (primarily wind and temperature)
• industrial and household emissions
• background and initial air quality
• transportation network
• fleet composition
• overall traffic load
• day of the week (including special days such as the beginning of summer vacations)
• special events (like major sporting events)
• strategic restrictions (speed limits, selective road closures for vehicle types, etc.);
• specific restrictions (construction, accidents, etc.).

Editing functions

An important aspect here is the translation of the more or less technical (and sometimes cryptic) model data requirements into concepts and terms that are directly problem relevant and directly understandable to the user. A general concept used is the specification of most user defined values in relative terms and as a selection from a list of predefined, valid and meaningful options. Relative specifications are, for example, changes in the overall traffic load or fleet composition expressed as percentage change relative to a or reference scenario.

The editing within SIMTRAP (at the level of the SIMTRAP Server) is supported by an embedded expert system that can be used to ensure

• completeness
• consistency
• plausibility

All parameters are represented by Descriptors which are terms used in the expert systems knowledge base. Descriptors are Objects that have several Methods available to determine or update their value in a given context (the scenario). One such method is to ask the user through an interactive dialog box.

A concrete example used for one of the relative (percentage) scaling factors would be:

DESCRIPTOR
scaling_factor
T S
U %
V very_small[  0, 10, 25]
V small     [ 25, 50, 75] 
V average   [ 75,100,125]
V high      [125,200,250]
V very_high [250,300,400]
Q What is the relative change, expressed in % of the
Q original value, that you want to apply to the 
Q for this scenario run ?
ENDDESCRIPTOR

In this example, the possible change is constrained between 0 (switching the value off completely) to 400, or a maximum of a fourfold increase.

Performance Variables

The performance variables measure the overall behavior of the system (in terms of a set of partly implicit and partly explicit objectives) in an aggregate form. This is clearly necessary for simple reasons of cognitive limitations: a scenario run of 18 hours, at 15 minutes output intervals, for a network with 2,000 links and a model domain of 10,000 cells, given only 5 link specific and 3 environmental parameters, will produce a total of 2,160,000 data items. For comprehension (as an elementary step toward comparative evaluation) they must be summarized in a few performance variables. These could include:

• compliance with environmental standards (yes/no) (for example, Council Directive 92/72/EEC on air pollution by ozone:
  • Health protection threshold
    0.110 mg/m3 for the mean value over eight hours
  • Vegetation protection threshold
    0.200 mg/m3 for the mean value over one hour
    0.065 mg/m3 for the mean value over 24 hours
  • Population information threshold
    0.180 mg/m3 for the mean value over one hour
  • Population warning threshold
    0.360 mg/m3 for the mean value over one hour
• average, maximum, and several spatio-temporal integrals of immission values,
• population exposure (spatio-temporally integrated)
• or an arbitrary spatially integrated (summed over n land parcels or model grids cells) environmental impact function of the type: where Ci represents the estimated immission in land parcel (grid cell i), Co is a reference or no-effects threshold, and Wi represents a landuse-dependent weight or penalty. The coefficients a and b describe the dose-effect behaviour of the pollutant in question.
• total traffic volume (car*km, person*km, tons*km)
• average speed, deviation from some reference/optimal speed
• average/total travel times, deviations from some reference optimal duration
• specific emissions (per person and ton*km).

Visualization

An additional important function provided by the user interface is the visualization of the scenario parameters. Due to the relatively large number (consider, for example, individual construction sites), graphical and symbolic representation is used to summarize numerous, and in particular spatially distributed, data.

In summary, simple scenario analysis results in a single (set of) result(s), that is (implicitly or explicitly) compared against a set of (absolute) objectives (expectations) and constraints such as environmental standards or some minimal requirements for average traffic flow.

Comparative Evaluation

Comparative evaluation requires that the performance variables of more than one scenario (minimally two for direct pairwise comparison) are displayed to the user simultaneously. For the spatially distributed (network or domain-grid specific data) this is accomplished by displaying equivalent data sets in two or four parallel windows. For the performance variables, this is accomplished by the parallel display, tabular and graphical, of the respective values.

In both cases, graphical and numerical, the side-by-side display can be augmented by the calculation and display of relative and absolute differences (deltas) of the respective scenario performance variables, for example, as a map of differential (increases and decreases) immission from two ozone concentration maps representing two separate traffic scenarios.

In summary, comparative scenario analysis results in direct comparison of two (or a set of) result(s), that are explicitly compared against each other and interpreted in terms of improvement or deterioration of performance variables vis a vis the objectives and constraints.

Discrete Multi-Criteria Optimization

Since each scenario is described by more than one performance variable or criterion, the direct comparison does not necessarily result in a clear ranking structure: improvements in some criteria may be offset by deterioration in others. This can only be resolved (and resulting in an eventual ranking and selection) through the introduction of a preference structure that defines the trade-offs between objectives.

The basic optimization problem can be formulates as:

As an overall decision support tool, we can now use a discrete multi-criteria approach to find an efficient strategy (scenario) that satisfies all the actors and stake holders involved in the traffic and environmental management decision processes.

The preferences of decision makers can conveniently be defined in terms of a reference point, that indicates one (arbitrary but preferred) location in the solution space. Normalizing the solution space in terms of achievement or degree of satisfying each of the criteria between nadir and utopia allows us to find the nearest available Pareto solution efficiently by a simple distance calculation.

Since decision and solution space are of relatively high dimensionality, the direct comparison of a larger number of alternatives becomes difficult in cognitive terms. The data sets describing the scenarios can be displayed in simple scattergrams, using a user defined set of criteria for the (normalized) axes. Along these axes, constraints in terms of minimal and maximal acceptable values of the performance variable in question can be set, leading to a screening and reduction of alternatives.

As an implicit reference point, the utopia point can be used. Consequently, and unless the user overrides this default by specifying and explicit reference point, the system always has a solution (the feasible alternative nearest to the reference point) that can be indicated and highlighted on the scattergrams and in a listing of named alternatives.

In parallel, graphical representation of the spatially distributed parameters can be shown as thematic maps. The visualization tools based on GIS and multi-media formats of the SIMTRAP server system supports a more intuitive and holistic understanding of alternatives that aids the definition of a reference point and thus supports the decision making process.