How AI is changing test automation: 5 examples

Don't get tripped up by thinking of the wrong kind of artificial intelligence (AI) when it comes to testing scenarios. It's less about HAL, the sentient computer from the movie 2001: A Space Odyssey, and more about statistics-based, machine-learning AI.

In fact, this second type of AI is already being used in some testing scenarios. But before looking at automation-testing examples affected by machine learning, you need to define what machine learning (ML) actually is. At its core, ML is a pattern-recognition technology—it uses patterns identified by your machine learning algorithms to predict future trends.

ML can consume tons of complex information and find patterns that are predictive, and then alert you to those differences. That’s why ML is so powerful.

Here are five test automation scenarios that already leverage AI, and how to use it in your testing successfully. 

World Quality Report 2018-19: The State of QA and Testing

1. Do visual, automated validation UI testing 

What kinds of patterns can ML recognize? One that is becoming more and more popular is image-based testing using automated visual validation tools. 

"Visual testing is a quality assurance activity that is meant to verify that the UI appears correctly to users," explained Adam Carmi, co-founder and CTO of Applitools, a dev-tools vendor. Many people confuse that with traditional, functional testing tools, which were designed to help you test the functionality of your application through the UI.

With visual testing, "we want to make sure that the UI itself looks right to the user and that each UI element appears in the right color, shape, position, and size," Carmi said. "We also want to ensure that it doesn't hide or overlap any other UI elements."

In fact, he added, many of these types of tests are so difficult to automate that they end up being manual tests. This makes them a perfect fit for AI testing.

By using ML-based visual validation tools, you can find differences that human testers would most likely miss. 

This has already changed the way I do automation testing. I can create a simple machine learning test that automatically detects all the visual bugs in my software. This helps validate the visual correctness of the application without me having to implicitly assert what I want it to check. Pretty cool!

2. Testing APIs 

Another ML change that affects how you do automation is the absence of a user interface to automate. Much of today's testing is back-end-related, not front-end-focused. 

In fact, in her TestTalks interview, "The Reality of Testing in an Artificial World," Angie Jones, an automation engineer at Twitter, mentioned that much of her recent work has relied heavily on API test automation to help her ML testing efforts. 

Jones went on to explain that in her testing automation, she focused on the machine learning algorithms. "And so the programming that I had to do was a lot different as well. … I had to do a lot of analytics within my test scripts, and I had to do a lot of API calls."

3. Running more automated tests that matter

How many times have you run your entire test suite due to a very small change in your application that you couldn't trace?

Not very strategic, is it? If you're doing continuous integration and continuous testing, you're probably already generating a wealth of data from your test runs. But who has time to go through it all to search for common patterns over time?

Wouldn't it be great if you could answer the classic testing question, "If I've made a change in this piece of code, what’s the minimum number of tests I should be able to run in order to figure out whether or not this change is good or bad?"

Many companies are using AI tools that do just that. Using ML, they can tell you with precision what the smallest number of tests is to test the piece of changed code.

The tools can also analyze your current test coverage and flag areas that have little coverage, or point out areas in your application that are at risk.

Geoff Meyer, a test engineer at Dell EMC, will talk about this in his upcoming session at the AI Summit Guild. He will tell the story of how his team members found themselves caught in the test-automation trap: They were unable to complete the test-failure triage from a preceding automated test run before the next testable build was released.

What they needed was insight into the pile of failures to determine which were new and which were duplicates. Their solution was to implement an ML algorithm that established a "fingerprint" of test case failures by correlating them with system and debug logs, so the algorithm could predict which failures were duplicates.

Once armed with this information, the team could focus its efforts on new test failures and come back to the others as time permitted, or not at all. "This is a really good example of a smart assistant enabling precision testing," Meyer said.

4. Spidering AI

The most popular AI automation area right now is using machine learning to automatically write tests for your application by spidering.

For example, you just need to point some of the newer AI/ML tools at your web app to automatically begin crawling the application.

As the tool is crawling, it also collects data having to do with features by taking screenshots, downloading the HTML of every page, measuring load times, and so forth. And it continues to run the same steps again and again.

So over time, it's building up a dataset and training your ML models for what the expected patterns of your application are.

When the tool runs, it compares its current state to all the known patterns it has already learned. If there is a deviation (for instance, a page that usually doesn't have JavaScript errors but now does), a visual difference, or a problem of running slower than average, the tool will flag that as a potential issue.

Some of these differences might be valid. For example, say there was a valid new UI change. In that case, a human with domain knowledge of the application still needs to go in and validate whether or not the issue(s) flagged by the ML algorithms are really bugs. 

Although this approach is still in its infancy, Oren Rubin, CEO and founder at machine learning tool vendor Testim, says he believes that "the future holds a great opportunity to use this method to also automatically author tests or parts of a test. The value I see in that is not just about the reduction of time you spend on authoring the test; I think it's going to help you a lot in understanding which parts of your application should be tested."

ML does the heavy lifting, but ultimately a human tester does the verification.

5. Creating more reliable automated tests

How often do your tests fail due to developers making changes to your application, such as renaming a field ID? It happens to me all the time.

But tools can use machine learning to automatically adjust to these changes. This makes tests more maintainable and reliable.

For example, current AI/ML testing tools can start learning about your application, understanding relationships between the parts of the document object model, and learning about changes throughout time.

Once such a tool starts learning and observing how the application changes, it can make decisions automatically at runtime as to what locators it should use to identify an element—all without you having to do anything.

And if your application keeps changing, it's no longer a problem because, with ML, the script can automatically adjust itself.

This was one of the main reasons Dan Belcher, co-founder of testing tool company Mabl, and his team developed an ML testing algorithm. In my recent interview with him he said, "Although Selenium is the most broadly used framework, the challenge with it is that it's pretty rigidly tied to the specific elements on the front end.

"Because of this, script flakiness can often arise when you make what seems like a pretty innocent change to a UI," he explained. "Unfortunately, in most cases these changes cause the test to fail due to it being unable to find the elements it needs to interact with. So one of the things that we did at the very beginning of creating Mabl was to develop a much smarter way of referring to front-end elements in our test automation so that those types of changes don't actually break your tests."

Become a domain model expert

Being able to train an ML algorithm requires that you come up with a testing model. This activity needs someone with domain knowledge; many automation engineers are getting involved with creating models to help with this development endeavor.

With this change, there is a need for folks who not only know how to automate, but who can also analyze and understand complex data structures, statistics, and algorithms.

Don’t panic! Keep automating

As you have seen, machine learning is not magic. AI is already here. Are you worried? Probably. Are you out of a job? Probably not. So stop worrying and do what you do best: Keep automating.

For more on how AI is changing testing, visit Joe Colantonio's AI Summit Guild online conference on May 30. 

Topics: Dev & Test