Firing up the Erudine engine
Building with behaviour
This is the next article in our occasional series on new, more formal (mathematically-based) approaches to system development. The first article (here), looked at Bayesian analysis and formal methods (which are only "new" to the general development space, of course).
This article looks at a tool, the Erudine executable behaviour engine, which uses the mathematical modelling of behaviour as a basis for systems development. Unfortunately (or, perhaps, fortunately), Erudine keeps its math a secret, so I can't say much about it – it usually sells by solving intractable development problems in a proof-of-concept, so one can probably assume that the math works.
Erudine automates the elucidation of system behaviours (expressed in the form of rules) and builds up a model of a system, which incorporates automatic consistency and completeness testing, and which can be executed in production. It suits problems with well-defined inputs and outputs (rather than algorithmic problems, although algorithms can be plugged in to the model).
It doesn't do automated Rule Induction - it's all human-steered. However, it can simplify rules for input to rule induction software and it's useful for pre and post-massaging of the inputs to and outputs from other systems.
From a developer's point-of-view, Erudine claims to be particularly useful for legacy replacement; even if the legacy application is undocumented and its programmers are no longer around. This is because it models the behaviour of the legacy application as a "black box" and completely reproduces this behaviour, as validated by current domain experts, on a new platform. Legacy replacement is a known "pain point" (where traditional approaches often don't work effectively) and is therefore a good place to get new technology deployment accepted. However, the Erudine approach should work on new developments too.
Erudine development follows a reasonably straightforward process:
Stage 1: This is a basic RUP (Rational Unified Process) inception process with domain experts supplying the input – even outside Erudine, this can be quite "agile", see here. Access to real business experts at this stage is important.
Stage 2: At this stage, the developers decide on the partitioning - morphology - of the system. An architectural framework is being built.
Stage 3: A significant piece of the system is chosen - a particular case (not "use case" exactly, it's all data driven) or set of dataflows. The scope of the Erudine model is defined in terms of physical input/output docs; reference data; or (at last resort) by hooking into the system and identifying the outputs from key dataflows. This stage delivers the "mental model" of the problem domain inside developer's heads together with a semantic glossary (usually, just a simple document) for use with the model.
This is where the basic computer model of the system dataflows is built; the developers pull out meanings and put placeholders in the model for scenario-based tests. Conceptual graphs are used to represent real system or domain knowledge – Erudine does this better than most other approaches can.
Stage 4: Now the developers add "behaviour" to the system, module by module – attacking the "low hanging fruit" first. They start with the bits that are easy to replace - or those that the client desperately wants to replace. They need to manage the delivery gap – the period when they're asking questions and nothing is being delivered. However, once delivery starts it is rapid because the most likely cases are being handled first. The end users of the system should also buy into the process early on, because of the focus on business behaviours and the involvement of domain experts from inception.
As each piece of behaviour is added, inconsistencies with early behaviours are highlighted and addressed – which is a form of continuous and automated unit testing.
Stage 4A: The complete model can be executed – the integration process – there is no need for a translation into code or switch over to conventional coding.
Stage 5: The model is run in order to test the behaviours it has captured (this is a combination of integration and user acceptance testing) and the system "plumbing" completed.
Stage 6: It now just remains to choose from various deployment options:
- If a greenfield site or new development, you can deploy the Erudine model as a conventional new system – cut over to the tested system from existing, possibly manual, processes (if any) once you're sure it's ready. However, Erudine seems not to have much actual experience of this.
- You can run the Erudine model in parallel with the original system and then cut over once everyone is comfortable with it. This is the safe, normative, option but it can be difficult to manage duplicate information or system state during the parallel run. This is an expensive option in the short term) and assumes a mature, well-organised company.
- You can use the Erudine model as "requirements doc" for a cheap rewrite using conventional coding techniques (probably using outsourced, cheap, programmers). This is the comfortable option and appears to minimise risk, but you are unnecessarily duplicating effort and forgoing the maintenance benefits Erudine promises (although resurrecting the Erudine model for maintenance will be cheaper than the initial build). If you rewrite the Erudine model you are probably writing new legacy. It's not really risk-free even in the short-term: it takes longer and you need to manage the rebuild quality and ensure that the Erudine behaviour isn't compromised in the rebuild.
A typical Erudine sell is based on identifying a pain point, such as a legacy system which must be replaced for good business reasons and for which a conventional rebuild is infeasible (or has actually failed). Generally, the sell goes:
- First, build new system related to the legacy target (important enough to matter; not so important as to be a company-killer) as proof of concept; then
- Second contract, recreate the whole legacy system behaviour with Erudine; and
- Potentially, maintain the system by changing its behaviour in the self-testing Erudine models.
As with any new approach, there are issues to consider. The Erudine approach to legacy reclamation is impressive, more so than the superficially attractive "put an object wrapper around your legacy and deploy it as a service" approach (which reads well but has serious problems in the detail – building standards-based chaos for one; and whether the legacy does break neatly into cohesive services for another). However, it may not be the only feasible option.
Micro Focus and others have mature automated tools for refactoring and understanding legacy systems by analysing the source code (if you still have it). Micro Focus, of course, has positive case studies (eg, one from as long ago as 2000, here). And, Compuware, for example, provides 4 GL tools such as Uniface for rapidly recreating legacy systems in a more agile, business-oriented environment. Compuware's approach may be an unfashionable one, but I think it's also a workable one.
Erudine offers a different approach and probably a more integrated one, with an attractive maintenance story. What makes it different is the underlying mathematical model of behaviour and consequent automatic consistency checking (as I said, this is very hard to assess objectively, as it is a secret "black box"); and the fact that it explicitly manages "knowledge" with conceptual graphs.
If it delivers on its promises, it moves rule-based systems to a higher level of knowledge management (the rules community was originally AI focused, now it plays down AI, perhaps Erudine puts it back, to an extent). Nevertheless, Erudine is difficult to evaluate without case studies – and case studies could succeed or fail for reasons unconnected with the use of Erudine. In any case, most of Erudine's best customers aren't talking publicly. Nevertheless, the secondary evidence of Erudine's workability from Gartner etc and from unattributable sources appears good. ®