ktsdesign - stock.adobe.com
For organizations creating algorithms and implementing systems, choosing between rules-based vs. machine learning-based systems is critical to the usability, compatibility and lifecycle of the application.
Getting outputs from a rules-based system can be a simple and nearly immediate application of AI, but an investment in machine learning can handle complex tasks with great speed. Enterprises must understand the core differences between the two, their individual benefits and the limitations of both before taking advantage of either.
Rules-based vs. machine learning
At the core of these two examples of AI, the logic and rules on which the systems or algorithms operate is what differentiates them. For rules-based systems, the logic that the system operates on is instilled at the beginning with little flexibility once deployed. First, a list of rules is created, often by an in-house developer, then an inference engine or semantic reasoner performs a match-resolve-act cycle, measuring information that it takes in against these rules. Here, human knowledge is encoded as rules in if-then statements for a specific rule. For example, in a rules-based algorithm or platform, a bank customer's personal and financial information can be measured against a programmed set of levels, and if the numbers were to match, then a home loan would be granted.
"In this scenario, the customer's information passes through an analysis process created by a human and built upon business rules provided to the developers," said Gus Walker, senior director at AI tech company Veritone, based in Costa Mesa, Calif.
On the other end of the spectrum lies machine learning, where rather than the algorithm being hand-coded by a human, it's created by selecting an appropriate AI model and presenting it with a very large dataset, Walker said. The algorithm then analyzes the dataset and determines relationships within that data; logic is embedded in the algorithm and was not coded by a human. As referenced in the name, the model trains itself and learns from the data, creating a cohesive relationship between data inferences and future data outputs.
Applying machine learning and rules-based systems
With different learning methods, deploying rule-based vs. machine learning systems is dependent on organizational need. Different types of artificial intelligence create different types of action, analysis or insight.
How to choose rule-based systems vs. machine learning is mostly dependent on how strict you want parameters to be, and who you'd like to do the learning -- a data science team, or an algorithm.
Machine learning is geared to handle complex and intensive issues with a relatively variable environment, while a rule-based AI system eschews black box training complications. However, the adaptability and speed of machine learning systems comes at a cost. There is an advanced level of commitment with machine learning, since in order to be properly trained, the algorithm requires a lot of data -- hundreds of thousands of records -- in order to make accurate predictions.
"It's only applicable for high-volume use cases, such as sales lead qualification or customer support auto-responses," Jeff Grisenthwaite, vice president of products at Catalytic, said. "Machine learning is also better suited to situations that have a large number of factors."
Training on this data can take extended periods of time, so when organizations have simpler tasks, taking a rules-based approach may make more sense.
"Rules-based systems are best suited to situations in which there are lower volumes of data and the rules are relatively simple," Grisenthwaite said. "Many companies use rules-based systems for expense approvals, defining the dollar thresholds that require management approvals at various levels."
The limitations and getting it off the ground
Though rule-based systems are typically not complex by nature, as more rules and mitigating factors are added in, it can become nearly impossible to add new rules without overlap. Taking the example of average income needed to be approved for a home loan, a rules-based system would have data scientists create a rule, such as income = 3 x estimated monthly mortgage payment for approval. However, it's never that simple in large AI applications; more modifications are needed and the nature of rule-based systems can become complicated.
These modifications can include anything from adding in locational complexities to factoring in economic downturn. These new rules significantly increase the amount of time and planning than originally anticipated. This increase in foundational work allows for more adaptability but takes more time. Therefore, mitigating the limiting factors can also decrease its ease of use.
Turning to machine learning models makes sense when the need for modifications in a rules-based system becomes overbearing. Creating an algorithm, training it on relevant datasets and teaching it to look for things takes more initial investment but can handle more complexity than a rules-based system. The model itself uses backpropagation for optimization to find the rule for approval (income = 3 x monthly mortgage x 0.025 credit score + 0.001 location) as it learns from data.
Their limitation lies in their reliance on quality data with which to train. The simple formula is that the quality of the data directly affects the quality of the model. The statistical models can take in many variables but needs to be trained on historical and relevant data. This requires similar but more likely greater work from relevant employees to find relevant data for the model to learn from.
If your model is attempting to discover new or rare events, as it would in fraud protection, the field from which examples can be harvested will be limited to start with. Not having enough labeled data will hinder the model's performance. Machine learning also requires an added level of expertise from your team. Writing the algorithm, identifying useful data and monitoring it all warrants a data science team and properly trained employees.