EVOTER: Evolution of Transparent Explainable Rule-sets

TL;DR

The framework introduces three innovations: time-lag operators, feature-feature comparisons, and nonlinear transformations that enable models that are both effective and auditable and advances the state of explainable artificial intelligence.

Abstract

Most AI systems are black boxes generating reasonable outputs for given inputs. Some domains, however, have explainability and trustworthiness requirements that cannot be directly met by these approaches. Various methods have therefore been developed to interpret black-box models after training. This paper advocates an alternative approach where the models are transparent and explainable to begin with. This approach, EVOTER, evolves rule-sets based on simple logical expressions. The approach is evaluated in several prediction/classification and prescription/policy search domains with and without a surrogate. It is shown to discover meaningful rule sets that perform similarly to black-box models. The rules can provide insight into the domain, and make biases hidden in the data explicit. It may also be possible to edit them directly to remove biases and add constraints. EVOTER thus forms a promising foundation for building trustworthy AI systems for real-world applications in the future.

Figures (16)

• Figure 1: Rule-set Representation in EVOTER ($a$) The BNF grammar for the rules, including time lags, feature comparisons, and feature exponentiation. ($b$) An example rule illustrating these concepts, with elements color-coded as in the grammar. Rules can have multiple conditions and a rule set consists of multiple such rules (not shown). In addition, EVOTER keeps track of how many times the antecedent is satisfied, and the rule set includes a default rule that is applied if none of the antecedents apply in the current situation.
• Figure 2: The three crossover strategies in EVOTER. ($a$) In single-point crossover, the offspring inherits rules from $parent_1$ up to its crossover point, followed by rules from $parent_2$ after its crossover point. ($b$) In uniform crossover, each rule is selected uniformly randomly from either parent at each point in the rule set; in addition, all remaining rules are included from the larger set. ($c$) In combine-rules crossover, the offspring is initially cloned from $parent_2$; a rule with a matching action in $parent_1$ and $parent_2$ is then selected, and its conditions are augmented with conditions from $parent_1$. Each of these crossover strategies serves a different purpose, and they are selected randomly during each crossover.
• Figure 3: Examples of permitted and rejected rule mutations in EVOTER. Gray highlighting on the right specifies the proposed mutation, green highlighting on the left identifies the results of permitted mutations, and red highlighting identifies the rejected mutation. In this manner, mutations that lead to contradictions and redundancies are automatically ruled out, making evolution more efficient.
• Figure 4: A sample rule set evolved to predict blood pressure. All the features are extracted from aggregations of MAP bp-sepsis. EVOTER discovered sets of features at specific time points to provide a useful signal for prediction. For instance, Std[4] specifies the standard deviation of the aggregated mean arterial pressure (MAP) over four minutes earlier. The evolved rules predict sepsis accurately and are interpretable and meaningful to experts.
• Figure 5: A sample solution rule set discovered for the Flappy Bird domain by direct evolution. The rules primarily identify situations where a flap is warranted. Rule 3 is redundant, and the conditions in all rules have redundancy, which is common in evolved solutions, making the search robust.