The incorrect or incomplete configuration of the Adaptive Decision Manager (ADM)
service causes many of the problems and technical issues related to adaptive models. Use
the following general guidelines to investigate and solve the most common
problems.
In the header of Dev Studio, click ConfigureDecisioningInfrastructureServicesAdaptive Decision Manager.
Ensure that there is an ADM node with a status of
NORMAL. If there is no node,
add one.
Clicking on the node will show additional metrics around total
number of model updates, number of models currently updating and
models waiting update and some more metrics around the update
times.
Click Edit settings, and then ensure
that the ADM settings are correct.
The ADM node contains settings related to monitoring. These
settings control the data that is written to the ADM data mart
and the schedule for ADM data monitoring.ADM settings
Enable extra logging for ADM to get detailed logs when you start and
stop the ADM node.
In the header of Dev Studio, click ConfigureSystemOperationsLogs.
Click Logging Level Settings.
In the Logger name field, enter
com.pega.decision.adm.
Set the Current level to
All, and then close the dialog
box.
Click Log Files to view or download the
current log files.
Note: Enabling logging at the whole ADM package level provides the most
comprehensive information to help you trace the root cause of the
issue. If the log files are too extensive, you can configure a
reduced logging scope.
Ensure that the pxHandleResponses real-time data
flow is running.
This data flow is not part of the core ADM, but if this data flow is
not running, then predictions that use a response timeout will not
emit a negative response (NoResponse)
when the specified waiting time expires.
In one-to-one customer engagement use cases, where Pega Customer Decision Hub is used, the data flow is started by the
PegaMKT-Engine agent schedule. The
pxHandleResponses data flow is a managed
run that should always be running.
Open the adaptive model rule, check the settings, possible outcomes,
and context identifiers.
Open the pxDecisionResults data
set.
The common predictor values (ADM inputs) and model executions are
stored in the pxDecisionResults data set. You can
browse or retrieve the results by interaction ID and subject ID. Each
decision result may reference
(pzModelExecutionReferences) one or more model
executions. An incoming response to a decision result updates associated
model executions if the pyOutcome field in the
response matches one of the allowed outcomes as defined in the adaptive
model rule.
In the header of Dev Studio, click ConfigureDecisioningModel Management.
In the Latest responses section, check
whether any responses have been sent to a model.
The Model Management landing page contains information around
the models created by ADM, positive and negative responses sent
to an ADM model, and the processed and recorded responses. For
more information, see Adaptive model details.
Expand the row for a response to view the details.
There is a separate page for each model execution that the
decision result is referencing. This page contains all the
information contained in the model execution, for example,
parameter values and context key values, and indicates
whether the model execution is
Updated,
Ignored, or
Monitored. An adaptive
model execution can be Ignored
if the outcome does not match any of the possible outcomes
defined in the adaptive model rule.
Model executions are also used for predictive models and
predictions, which use them for monitoring.
Check which models are executed in the decision strategy. Use the
strategy test panel and the Executed models under
internal fields to track which model executions are propagated.
The Group By and Data
Join components both have options that control which
model executions are propagated.
Capture response flow and set response strategy.
Responses can be captured by interaction ID (and subject ID) or by
past period (only subject ID). The strategy requires an external input
shape.
Previous topic
Troubleshooting the Adaptive Decision Manager service
Next topic
Adaptive models do not contain any active predictors, or contain only one active predictor
