MilramX Logo

Strengths and Weaknesses of Decision Trees
 for Coding Real-Time Artificial Intelligence Applications

MilramX Information Transfer

Some of the earliest real-time AI systems were based on decision trees, where inputs to the system were used to traverse a fixed tree with conditional rules applied at each level of the tree.

These are still used where the size of the decision tree is limited, the outcome of each decision process can be expressed in a yes/no decision, and the decision flow can be hard coded, such as with a limited rule-set within a MilramX intelligent agent.

Applications today include ladder-logic programming of PLCs (Programmable Logic Controllers) as well as SCADA (Supervisory Control and Data Acquisition) systems used in process control and some algorithms within self-guided cars.

The decision tree paradigm breaks down, as the code becomes impossible to manage, debug or easily modify when:

  1. The Decision Tree becomes too large to handle in one body of code - too many decision processes and too many branches with too many levels.
  2. The output of each decision process is not a yes/no but a set of answers with different probabilities. This occurs where the decision process is some form of statistical correlator, such as a neural network.
  3. The decision flow cannot be expressed in a hard-coded tree hierarchy. Here data needs to flow between different decision processes depending on the outcome of the analysis, in the form of a connected graph, rather than a tree.

These days, building an AI system is an iterative process, which requires adding and changing decision rules and processes in an "Agile" software development process.

Except in simple cases, such as within intelligent agents, decision trees do not lend themselves to this programming style and are much more effective with the older "Waterfall" development methodology where all the decisions are hard-coded based on a fixed specification.

Return to "Why Intelligent Agents?"