Page 288 - Artificial Intelligence for the Internet of Everything

266 Artificial Intelligence for the Internet of Everything

to update its parameters, especially when its estimate does not match the final
decision. In addition to catching errors, this phase is also crucial for building
human trust in the system.
In the third phase, the AI model has been validated sufficiently to allow it
to control the settings directly, having performed successfully for some time
in the second phase. We still send the model’s decisions to the manager for
oversight, and report any errors back to the model, but this should now be
regarded as an exceptional situation rather than the norm.
In most cases, of course, HVAC system settings are low-stakes decisions
and errors are not too costly, but the same process logic applies when we
introduce nearly any AI system into existing human-driven processes. For
more safety-critical systems, we can use the same sort of contract semantics
developed in the previous section to describe process requirements.
Inparticular,noticethatallthreediagramshavethesameinputsandoutputs;
since they share the same interface, they may participate in the same contracts.
Thus,atahigherlevel,wemightplacerestrictionsontheoverallprocess,suchas
limiting the overall energy costs or mandating certain situational responses, and
wecananalyzeeachofthethreephasestoensurethattheymeetthesegoals.This
ensures that we can modify the low-level process without impacting higher-
level performance. With more sophisticated semantics bringing in probability
or temporality we could also begin to analyze questions like when we should
shift from one phase to the next.
This gives an indication of some of the ways that better system represen-
tations arising in CT can support the application of AI in many contexts. We
can use AI to address object-level problems of the system, thinking of AI
models as just another component. At the same time, we can use AI to
address metalevel problems of the system, assisting in design and decompo-
sition decisions. We can use CT representations to document AI algorithms,
both internally and in their interaction with other system processes. By pro-
viding a precise, technical language for expressing and analyzing system rela-
tionships, CT can help to structure the use of AI across a wide range of
contexts and applications.

13.8 CONCLUSION

In this chapter we have argued that new, more expressive representations are
needed to support the design, engineering, and analysis of modern complex
systems, and we have suggested that the mathematical field of CT can pro-
vide them. We have seen how CT provides a uniform approach to defining