Future Work - Perfect Failure Masking Phenomenon

[christianadriano]

Current assumption is that if there an observation that indicates a failure on a shop, the shop might or not have a failure. However, if there are no observations indication failures on a given shop, then, we assume currently, that there no failures on that shop. In other words, we assume that absence of evidence is evidence of absence!

In the future we might want to relax this assumption (usually deemed to be too strong in real world). For that we would need to investigate the latent failures, for which we do not have any observation (no indication) of failure on a given shop. This latent failures might happen because of phenomenon of perfect failure masking, where one failure ocludes the other

• How to produce the phenonmenon of failure masking?
• How to detect pefect failure masking?
• How to find the latent failure?
• How to mitigate the effect of failure masking on the agent side?

[jocodeone]

For enabling the agent to handle a perfect failure masking, the action space has to include "NOOP" as an option that implies that the agent won't do anything.

For detecting a perfect failure masking I see several options:

• Change the NN to include RNN in order to include the last steps in the decision of whether this observation does not include any failure
• The network could have several inputs. At the moment only the component state is used to choose an action. However, the current utility and all the other values not taken into account could be important to decide whether a failure has occurred or not