IBM Cloud Docs
Algorithm version and training

Algorithm version and training

You can use the Algorithm version setting to choose which watsonx Assistant algorithm is used for training. Your assistant is trained when you update the content or settings, or through automatic retraining.

There are three choices:

  • Beta: Use Beta to preview and test what is coming. The capability in the beta version is likely to become a supported version later on. It's not recommended to use the beta version in a production deployment.
  • Latest: The current supported version that's recommended for your live production assistant.
  • Previous: The last supported before the latest version. Support for this version ends when the next version is released.

To choose an algorithm version for actions:

  1. On the Actions page, click Global settings Gear icon.

  2. Click the Algorithm Version tab.

  3. Choose a version, then click Save.

To choose an algorithm version for dialog:

  1. On the Dialog page, open the Options section.

  2. Click Algorithm Version.

  3. Choose a version.

The latest and previous versions have date labels such as Latest (15-Apr-2023) or Previous (20-Dec-2022). See the watsonx Assistant release notes for details about each algorithm version release.

Algorithm version choices are currently available for Arabic, Chinese (Simplified), Chinese (Traditional), Czech, Dutch, English, French, German, Japanese, Korean, Italian, Portuguese, and Spanish. The universal language model uses a default algorithm.

Automatic retraining

IBM Cloud

IBM® watsonx™ Assistant was released as a service in July 2016. Since then, users have been creating and updating skills to meet their virtual assistant needs. Behind the scenes, watsonx Assistant creates machine learning (ML) models to perform various tasks on the user's behalf.

The primary ML models deal with action recognition, intent classification, and entity detection. For example, the model might detect what a customer intends when they say I want to open a checking account, and what type of account the customer is talking about.

These ML models rely on a sophisticated infrastructure. There are many intricate components that are responsible for analyzing what the user says, breaking down the user's input, and processing it so the ML model can more easily predict what the user is asking.

Since watsonx Assistant was first released, the product team has made continuous updates to the algorithms that generate these sophisticated ML models. Older models continued to function while running in the context of newer algorithms. Historically, the behavior of these existing ML models did not change unless the skill was updated, at which point the skill was retrained and a new model that is generated to replace the older one. This meant that many older models never benefited from improvements in our ML algorithms.

IBM® watsonx™ Assistant uses continuous retraining. The watsonx Assistant service continually monitors all ML models, and automatically retrains those models that have not been retrained in the previous 6 months. Your assistant retrains by using the selected algorithm version. If the version you selected is no longer supported, watsonx Assistant retrains by using the version that is labeled as Previous. This means that your assistant automatically has the supported technologies that are applied. Assistants that were modified during the previous 6 months are not affected.

Usually, this retraining is seamless from a customer point of view. The same inputs result in the same actions, intents, and entities being detected. Sometimes, the retraining might cause changes in accuracy.

Instructions for watsonx Assistant for IBM Cloud Pak for Data

IBM Cloud Pak for Data

When you upgrade your instance of watsonx Assistant for IBM Cloud Pak for Data, if your existing models were trained by using an algorithm version that is still supported, your models are not retrained during or after the upgrade.

New algorithm versions are included in major releases (for example, 4.0.0 or 5.0.0) or minor releases (for example, 4.5.0 or 4.6.0). A monthly release might include a new algorithm version if there is more than 6 months between major or minor releases. Each algorithm version supports 2 major or minor releases or a maximum time of 12 months, whichever is first. For more information, see the IBM Cloud Pak for Data Software Support Lifecycle Addendum.

Each new release includes full support for the version that is listed as Latest in the most recent prior release. This version is then labeled as Previous after you upgrade. In addition, each new release supports running models that were trained on the version before that so that upgrading doesn't affect your runtime. For example, if you upgrade from IBM Cloud Pak for Data 4.6 to 4.7, and were using Latest (01-Jun-2022) that version becomes listed as Previous (01 June 2022) and remains your selected version.

Automatic retraining after you upgrade

After your watsonx Assistant for IBM Cloud Pak for Data upgrade is complete, watsonx Assistant performs automatic retraining for any assistant models that were trained by using a version that is no longer supported. In this case, watsonx Assistant automatically retrains your assistant to the Latest version. This automatic retraining is required to ensure your ability to run your trained models in your next upgrade.

Best practices

It's recommended to use the Latest version in your production deployment of watsonx Assistant for IBM Cloud Pak for Data. This is the default for new assistants. During an upgrade, your settings don't automatically switch existing assistants to use the latest version. If before your upgrade you selected Latest, your settings continue to use that version, now labeled as Previous. After you upgrade, it's recommended you choose Latest and run basic regression tests.

IBM performs robust testing on various data sets to minimize impacts on existing assistants. But given the nature of machine learning models and the nuance and subtlety of natural language processing, you might find some discrepancies from version to version. If you find a major issue through your tests, you can switch your settings and use Previous to return to the previous behavior. In this event, we recommend you contact IBM and provide details of your test so that that IBM can support you in the steps to resolve the problem.

It's also recommended that you try the Beta version in one of your test systems after you upgrade. This gives you early visibility to changes that are likely to be delivered in a future version, and reduces the probability of negative impacts to your production systems. IBM values both positive and negative feedback from customers who use Beta. You will have the opportunity to shape how the algorithms function before the version is promoted to Latest in a future version. If you choose Beta, your assistant always trains on the most current beta version.