Excerpt from Enterprise Architecture Fundamentals:
While implementations, especially the Deep-learning variants, combine a wide range of established technologies (e.g., modal logic, game theory, nonparametric statistics), their main characteristic is their ability to extract implicit knowledge from raw data and morph it into symbolic representations.
These technologies already make up the nuts and bolts of business intelligence, but their broader significance for enterprise architecture is an open issue (cf. chapter 16). Based on game theory’s basic distinction between zero- and nonzero-sum situations, ML can be employed in three basic contexts (figure 9-11):
- One-sided (no identified parties) issues make up the bulk of ML applications, continuing the traditional operational research on the symbolic side of knowledge (KK), or defining new frontiers with pattern matching on the non-symbolic side of knowledge (KU).
- Multisided, zero-sum issues entail fully assessable outcomes, enabling improvements based on comparisons. Given symbolic representations, ML can rely on computation and reasoning (KK); otherwise, it relies on statistics (KU).
- Multisided, nonzero-sum issues result in open-ended and multidimensional outcomes that cannot be assessed unequivocally. With symbolic representations (KU), qualified assessments — and therefore learning — can be achieved with supervised Deep learning; otherwise (UU), Reinforcement learning is necessary.
If only to provide guidelines, that taxonomy can be aligned with EA concerns: operations and business intelligence (one-sided), systems (zero-sum), and enterprise (nonzero-sum). On that basis, the use of ML at the enterprise level is meant to be collective and detached from domain-specific concerns, with rationales and outcomes embodied in organization and systems. As a consequence, the issue of knowledge traceability (why is that true?) is compounded by one of organizational accountability (who contributed?). Both issues will be considered in the next chapter.