How can you integrate machine learning (#ML/#AI) into a rule-based system? (when the rules are already known and need not be inferred from data)?

Sometimes, people ask me simple but profound questions on AI

When I was in India (Nashik) at a social event - Mādhav Dābké asked me this question

To paraphrase

How can you integrate machine learning (ML) into a rule-based system? (when the rules are already known and need not be inferred from data)?

Madhav is a board member of GMG, mentor and former banker turned banking technology transformation consultant who currently heads banking products for a multi national company.

The response to this question is quite complex. Here is a partial response (with other blogs to follow as below)

Background

AI is all about finding hidden rules as the above (well known) simple diagram shows

In the simplest case, the rules are known and encoded (ex programming/software development)

In the next complex case, the rules are derived from the data (machine learning) based on request response pairs

You could also derive both the rules and the structure from the data(deep learning)

Now, there are some more complexities when it comes to rules

1) LLMS have a world view and hence have causality built into it (to some extent) based on the rules that they can infer

2) You can think of ?simulations as creating vast amounts of synthetic data that resemble real-world scenarios, especially when actual data are scarce. In other words, knowing the underlying rules, you can use simulations to generate data.

Both these require some more discussions (and separate blog posts).

Now coming to the question

How can you integrate machine learning (ML) into a rule-based system? (when the rules are already known and need not be inferred from data)

Here are the possibilities

Machine learning and rule based systems

The most obvious one is: ?

1) Enhance Rule Selection with ML Models i.e. Use ML to determine when and which rules to apply. Here, you can use historical data to train a classifier to predict which rule or set of rules is most applicable to the current situation ex: In a medical diagnosis system, use ML to analyze symptoms and select the most probable diagnostic rules based on patterns learned from past diagnoses.

2) You could also refine Rules with ML Insights by adjusting the parameters of the rules based on data

example: In a financial fraud detection system, adjust the thresholds for certain indicators of fraudulent behavior based on trends learned from historical data (using an LLM model)

3) You could also scale by Preprocessing Data for Rules both for inputs and outputs

ex Use clustering to group similar financial transactions and apply different sets of rules to each cluster. (scaling inputs) and

After applying rules to classify texts into categories, use an ML model to refine these classifications based on additional text features not considered by the rules. (scaling outputs)

4) Next, you could use Anomaly Detection to Inform Rules ex: In network security, use anomaly detection to identify new patterns of intrusion that are not covered by existing security rules.

Then there are some more sophisticated possibilities

a) ?Use Reinforcement learning to select rules dynamically

b) ?Hybrid Rule/ML Models for rule generation

c) ?Use LLMs to train the rules to generate the output?

The workflow is a bit different also Essentially

1) Identify the Limitations: Understand where the rule-based system falls short (e.g., handling new situations, needing manual updates).

2) Collect and Prepare Data about the rules themselves: Gather historical data that includes the situations, rules applied, and outcomes.

3) Choose one or more of the above strategies to train the model

welcome thoughts

If you want to be a low code data scientist see our course on low code data science at the university of oxford?

If you are interested in AI and Digital twins please see University of Oxford course on AI and Digital twins

see also

An AI first educational institute for empowering domain experts(non coders) to learn artificial intelligence (AI)

with

Garje Marathi Global Ravindra Gadgil Anand Ganu Abhijit Deshpande Ashutosh Deshpande