The Single Value Monster

Most programs that do calculations are limited to the transmission of a single value. An algorithm knows how to calculate a value, the resulting value is stored at a location in a data structure. If multiple possible values were to result from the calculation, there would be nowhere to store them, and no clue as to when the calculation should be revisited if the number of possibilities might have reduced in the meantime.

The algorithm - single value approach, found in spreadsheets and project management applications, fits well with particular tiers in an organisation - what might be described as the "fingers and toes", where an immediate stereotyped response to an input is desired - the  "reflex arc". A call centre handling a mortgage application based on a form doesn't need much subtlety or flexibility of response, but if the call centre widens its reach to include financial advice....

Complex transactions between businesses are one area where single values, at least initially, in the transaction are not helpful to a resolution. The transaction depends on what each party can bring to it, and order of magnitude rather than precision is important -

"If you could buy more than 50,000, we could get the price below $2.75."

In reality, there are many more factors than quantity, price and delivery, and each one has some influence on the others. The transaction is more like settling of ranges of possibilities on both sides of the transaction to a common and mutually acceptable state.

Organisations which buy and sell risk, insurance companies for instance, may need to see the "big picture" when handling transactions that could change their risk profile. Here again, the simple single value approach at each calculation will not allow the creation of a more complex view in the analysis.

There is another level in large organisations - the planning or strategic level, where a simple and immediate response is usually not possible. What is being handled is too complex, too dynamic to admit of a single possibility at each stage in the analysis. Alternatives need to be carried forward or decisions need to be rolled back, making the single value approach counterproductive in that it forces decisions at inappropriate points in the analysis.

A database is a good example where looking separately at each single record can blind you to the big picture. The database has tools (some version of SQL) to allow you to see clumps, to find where the weight of data is headed. Connecting the database, with its ability to handle sets, to a single value analytic tool like a spreadsheet destroys the versatility and flexibility of the analysis that is possible directly on the database. Even more, the database will often contain implicit data structures that span multiple records and are relatively opaque to SQL. An example is trips in an airline database. The information is there in the database, just difficult to extract. A premise of SQL is that all records are independent, but a trip links records together, making a sequence of records dependent on each other. Sequences of records may overlap each other, making them even more difficult to find and operate on.

An alternative to the algorithmic - single value approach is the active structure approach, where multiple alternative values can be transmitted through the network and the network connections provide the phasing for re-analysis when the number of alternatives may have changed (sometimes the alternatives increase, when a jump to another state, inconsistent with the first, has occurred).

If we are to extract relations and patterns from a database, and particularly wish to be sensitive to the higher level structures that may be present, it would seem reasonable to couple the database to an analytic structure that shares with the database analytic language the ability to bring together multiple sets of data.

Constraint Solvers appear to be able to handle sets of data, but the sets are built at compile time, and are not propagatable, the system merely cycling through the set and handling them as separate single values.

The Single Value Monster appears strongly in the area of statistics. Much effort is spent in obtaining a metric, such as mean or standard deviation, to stand as a model of some behaviour. Add a few of these statistical models together, and the accuracy is very poor. While computing was done by hand, this approach was necessary. Now that computers do the calculations, we can move away from single values towards simple models of behaviour using connections which capture far more of reality than the single value does. The simple model is easily constructed to connect to many sources, so influences flowing into the model can capture more diversity of behaviour.  We combine several simple models together, with no thought of direction, and we can easily handle very complex behaviour.

NLP is an area where single values, computed at the wrong time and with insufficient information, leads to extremely poor quality outcomes - see A Pipeline to Failure in Bioinformatics.

Technical Discussion