Genetic Algorithm Rule Labeling for Persistent State Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data mining techniques using genetic algorithms face challenges in domains with persistent states, where systems forget previous actions and decisions, leading to suboptimal performance in tasks like robot target pursuit and securities trading, due to lack of memory about the reason behind previous assertions.
Innovation Solution
Incorporating a label field in rule outputs to maintain the reason for previous assertions, allowing rules to evolve based on the state of the system, including the reason for previous actions, and using an experience-layered elitist pool to ensure competition among individuals with similar experience levels, promoting diversity and accuracy in fitness estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If genetic algorithms are used for data mining in domains with persistent states, then the system can process and learn from historical data, but the system loses memory of previous actions and decisions, leading to suboptimal performance
Solution Approach 1:
A label field is introduced as an intermediary component in the rule output structure. This label field carries forward information about the reason for previous assertions, serving as a mediator between past decisions and future actions. The label is passed through the system's memory structure, allowing the genetic algorithm to indirectly access historical decision context without requiring the system to directly remember previous actions.
Solution Approach 2:
The label field creates a feedback loop where information about previous assertions is fed back into the system's decision-making process. By including the label in the rule output and using it in subsequent rule evaluations, the system establishes a feedback mechanism that preserves knowledge of past actions, enabling more informed future decisions in domains with persistent states.
2Productivity
If the population converges to local optima, then the algorithm stabilizes and reduces computational overhead, but no further improvements are made in performance
Solution Approach 1:
The replacement strategy is made dynamic by introducing experience-based weighting. Instead of a static replacement approach, the system dynamically adjusts which individuals are replaced based on their experience levels and performance trajectories. This dynamic approach allows the population to escape local optima by strategically replacing individuals who may be trapped in suboptimal solutions while preserving those demonstrating consistent improvement.
Solution Approach 2:
The fitness evaluation parameters are changed to incorporate experience-based metrics. Rather than relying solely on raw fitness values, the system modifies the evaluation parameters to consider an individual's experience level and improvement trajectory. This parameter change allows the algorithm to distinguish between individuals that are merely stable versus those that are genuinely optimized, enabling continued improvement even when the population appears converged.
3Ease of operation
If individuals with different experience levels compete together, then the selection process is simplified, but competition is unfair and diversity is reduced
Solution Approach 1:
The population is segmented into experience layers based on the number of generations each individual has survived. This segmentation creates distinct competitive groups where individuals with similar experience levels compete against each other. The segmentation maintains selection process simplicity within each layer while ensuring fairness and preserving diversity across the entire population by preventing premature convergence dominated by early-surviving individuals.
Data Source
AI summary
Roughly described, individuals in both a training system and in a production system include a label field in their rule outputs. Positions entered by an individual are maintained in a status record for the individual, including the label output by the rule which triggered entry of that position. Rules that assert exiting or partial exiting of a position also output the label from the rule which triggered the assertion, and are effective only so far as matching positions exist or remain in the individual's status record, including a matching label. Labels present in the status record also can be referenced in conditions of a rule. During evolution, a rule's output label is subject to crossover and/or mutation just like the conditions and output assertions.