Achieve real-time fraud scoring through co-locating an AI model with the IBM Z applications managing your transactions. This enables you to analyze 100% of transactions running through your on-premises systems, and significantly reducing risk to both your business and customers.

Eliminate the manually intensive auditing process through a machine learning model trained in detecting fraud for your claims processing systems. This enables you to rapidly analyze claims and significantly reduce overall human intervention.

By training a model on risk exposure and co-locating it with transactional workloads, unlock insights about risk on a transaction-to-transaction basis.

Co-locate an inferencing model for analyzing loan applications, alongside the automation of rules-based business decisions, to achieve rapid insights, low latency, and minimal exposure of customer and lender data.

Co-locate an inferencing model for analyzing correlations between weather patterns and insurance risk alongside transactional systems, allowing you to achieve rapid insights.

Enable accurate analysis of satellite imagery with a deep learning learning model deployed to the IBM Z platform, ensuring the sensitive care of aerial images while driving rapid insights.

Achieve effective and energy-efficient computer vision for medical imaging through the training and inferencing of AI models on the IBM Z platform.

In order to successfully install and configure MLz, some additional IBM and open source technologies should be in place.

The minimally required technologies include:

• IBM z17, z16, z15™, z14, z13®, or zEnterprise® EC12 system
• z/OS 2.5 or 2.4
• z/OS UNIX System Services configured
• z/OS Integrated Cryptographic Service Facility (ICSF)
• z/OS OpenSSH
• IBM 64-bit SDK for z/OS Java Technology Edition Version 8 SR7, or Version 11.0.17 or later
• IBM WebSphere Application Server for z/OS Liberty version 22.0.0.9 or later
• Python AI Toolkit for IBM z/OS (for MLz Enterprise Edition)

Note: Additional prerequisites will be required if you plan for specific use cases, for example, to use the IBM z16 on-chip AI acceleration for scoring ONNX models or to run a scoring service in a CICS region. For a full listing of prerequisites broken down by use case, view Planning system capacity for ML for IBM z/OS: Basic system capacity.

MLz performs the best when adequate system capacity in terms of server, processor, memory, and disk space is available. Allocate sufficient capacity for MLz on your z/OS system to meet the demands of your enterprise machine learning workload.

Note that additional capacity may be needed depending on the training workload you plan to run on IBM Z. This may be influenced by the number of concurrent training jobs, the type of models being trained, the size of the training data set, the number of data features, and the specification of model parameters. For guidance on planning capacity for your training workloads, view Planning system capacity for ML for IBM z/OS: Capacity consideration for training workload on Z.

USS skills are necessary throughout the configuration process, including configuring essential user IDs, port definitions, and authentications for the product.

For a full list of which configuration steps require USS skills, view Roadmap for installing and configuring ML for z/OS Enterprise.

There are a number of terms you will encounter when configuring MLz that are essential for understanding how the product is set up.

For a list of essential terms to know, view Commonly used terms and definitions for installation and configuration.

Your security administrator can help you configure a MLz setup user ID, which is necessary for using the MLz configuration tool. For ease of access control, it is recommended that you plan to also use this same user ID for MLz services.

For further information about configuring user IDs, view Configuring user ID for setting up Machine Learning for IBM z/OS Enterprise.

Your security administrator will also be responsible for ensuring that a RACF KeyRing is created for user IDs that must be created for MLz, and will help to establish the requisite IDs and authorizations. They will also be involved in securing network communications for MLz with AT-TLS.

For further information about securing network connections and configuring the keystore, view Configuring secure network communications for MLz.

Your network administrator can help you reserve the various ports which need to be dedicated to the purpose of MLz communicating across systems and services. A number of required and optional ports should be accounted for prior to using the MLz configuration tool.

For further information about configuring ports for MLz, view Configuring ports for ML for IBM z/OS.

Your z/OS database administrator can help you plan for the storage of metadata objects in the repository of your choice: an embedded database with MLz or a Db2 for z/OS database.

For further information about creating and storing metadata objects, view Creating metadata objects for the repository service.

You must select one of two options for storing important metadata about models. Selecting the embedded MLz database doesn’t require further configuration, but limits the availability of MLz services. Selecting Db2 for z/OS allows for high availability, but requires that you have a ready Db2 system, and also requires that you provide network, authentication, and schema details for MLz metadata objects.

You must specify host system details for both the MLz UI and core services, as well as credentials for the default admin user and the MLz application developer. You can also enable audit trace for governance of model data, model versions, and model deployments. Lastly, if you had selected Db2 for z/OS as your metadata repository, you can add MLz core services to a cluster, which routes services to an alternate MLz instance if one isn’t available.

You must create a name for the MLz runtime environment, as well as provide details for both the Spark runtime engine and the Python runtime engine. The Spark runtime engine is a high-performance analytics engine for large-scale data processing able to perform in-memory computing, and you will need to allocate various ports for its configuration, as well as determine the maximum number of retries for port binding. You will also have the option of enabling client-side Spark authentication. The Python runtime environment allows for the use of a Python-based scoring service, as well as the enablement of a Jupyter Notebook server for model development in the MLz UI. For Python, you will need to specify the Python virtual environment name and the Python packages installation source.

You must provide a name for the MLz scoring service. A scoring service is where your AI models are deployed to for real-time inferencing. You can choose to create a standalone scoring instance, or to set it to a cluster for high availability. Creating a cluster scoring services requires that you had selected Db2 for z/OS as your metadata repository, and will require you selecting an available cluster to join, or creating a new one by providing network details for the sysplex distributor.

You must choose whether to enable Trustworthy AI features for your MLz instance. If enabled, you will need to provide network and authentication details for an IBM z/OS Container Extensions (zCX) instance to support the features.

You can import models to MLz directly through the UI. In the Models tab, click “Import model”, and then select the relevant tab for whether you want to import from file or from IBM Cloud Pak for Data.

If importing from file, you will need to provide a name for the file to be displayed in the MLz UI, and you will need to select the model type.

If importing from IBM Cloud Pak for Data, you will need to provide connection and authentication details, as well as provide a name for the file and select the model type.

You can create models in MLz with a Jupyter Notebook server which can be accessed directly from the UI. In the Models tab, click “Create model” to open a view of the Jupyter Notebook interface which you can use to create and save new models for MLz. If you’re unfamiliar with the interface or basic workflows of notebooks, view the open source documentation published by Project Jupyter.

When utilizing machine learning and deep learning with your transactional applications, it’s important to continually ensure that deployed models are working as intended. MLz offers a method for evaluating the performance of SparkML and PMML models to ensure accuracy.

To do so, select a model from the Models tab, open the overflow menu, and click “Create evaluation”. Then add a data source and choose whether to run the evaluation over a time range, and click “Create”. The evaluation for deployments of that model can then be triggered using the overflow menu in the Deployments tab. The evaluation can also be scheduled to run repeatedly.

For full detail about this procedure, view Evaluating and reevaluating a model.

A new dashboard for generating and managing SHAP and LIME explainers is available in MLz Enterprise Edition v.3.1. You can generate visualized local explanations of specific AI predictions derived from a given model deployment.

To do so, click “Explanations” in the global navigation, or click “Explain” from the action menu of a given model deployment in the Deployments table. Then you will create a subscription, which associates the model deployment with available services such as LIME and SHAP explanations. Upon filling in the necessary details and uploading training data, you can select the explainer type you wish to use in the subscription. Then you can filter the transactions available within the deployment to target a specific subset you want to make available for explanation.

Once the subscription is ready, you can search and/or filter for the transactions for which you want to generate local explanations. You can then create the explanations. Once loaded, will provide visualized results of feature importance using LIME and/or SHAP (depending on your selections during subscription creation).

For full detail about the available trustworthy AI features, view Trustworthy AI.

For CICS, the MLz scoring engine can be hosted in a WebSphere Liberty server within the CICS runtime and provides a program ALNSCORE that can be invoked with EXEC CICS LINK, passing the data using CICS channels and containers.

For further detail, view Planning AI infusion into applications on IBM Z.

For IMS, the MLz WOLA (WebSphere Optimized Local Adapter) scoring interface can be invoked from IMS COBOL applications for high performance real-time inferencing in IMS online transactions.

Refer to the Deploy AI models for real-time inferencing article for further detail.

For WebSphere, a Java API can be used to call the MLz scoring feature configured in a WebSphere server.

For further detail, view Planning AI infusion into applications on IBM Z.

An ODM rule driven by the runtime application can be enhanced to reference a model deployed to MLz, and then use the prediction from the model in the rule. ODM uses a highly efficient interface between ODM and MLz.

Many CICS and IMS applications already use ODM rules to inform their decisions, in which case a rule called by the application can be enhanced with machine learning to drive an AI model deployed to MLz. If the application does not currently use ODM rules, it can be updated to use an ODM rule that drives the AI model via MLz, and hence include additional insight in the result from the rule.

For further detail, view Planning AI infusion into applications on IBM Z.