XCS for Sequential Perceptual Aliasing in Multi-Step Decision Making

Area of Science:

• Artificial Intelligence
• Robotics
• Cognitive Science

Background:

• Sequential perceptual aliasing poses a significant challenge for learning agents, hindering their ability to differentiate states and make optimal decisions.
• Current systems often struggle with effective abstraction and discrimination of observations, limiting policy learning.
• This necessitates novel approaches to handle complex state representations in robotic systems.

Purpose of the Study:

• To introduce new types of sequential aliasing and propose an enhanced XCS classifier system.
• To enable learning agents to effectively manage and learn from aliased states in sequential decision-making tasks.
• To improve the accuracy and efficiency of policy learning in the presence of perceptual aliasing.

Main Methods:

• Introduction of new aliasing types within the sequential aliasing context.
• Development of a hierarchical Frames-of-References-based XCS (Hi-FoRsXCS) classifier.
• Implementation of a complete state-action map for learning.
• Concatenation of sequences of aliased states with identical observations into a chain.

Main Results:

• The proposed Hi-FoRsXCS system successfully chains sequences of aliased states.
• Hi-FoRsXCS predicts associations between observations and aliased states using the ends of the state chain.
• Experimental results show that Hi-FoRsXCS significantly outperforms existing systems in terms of accuracy.
• The system enables optimal policy learning with a complete action map.

Conclusions:

• Hi-FoRsXCS offers a robust solution for sequential perceptual aliasing in learning agents.
• The chaining mechanism effectively addresses the challenge of differentiating states with similar observations.
• The enhanced system demonstrates superior performance in policy learning accuracy compared to prior methods.
• Further discussion on the limitations of Hi-FoRsXCS is provided.