A planar tracking strategy based on multiple-interpretable improved PPO algorithm with few-shot technique.

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Machine Learning

Background:

• Planar tracking problems require sophisticated algorithms.
• Traditional Proximal Policy Optimization (PPO) can be sample-inefficient and lack interpretability.
• Real-world interaction data is often scarce and costly to obtain.

Purpose of the Study:

• To propose a novel algorithm, F-GBQ-PPO, that enhances interpretability and reduces sample complexity for planar tracking.
• To investigate methods for increasing the comprehensibility of tracking policies.
• To address the challenge of limited real interaction samples in practical applications.

Main Methods:

• Introduced a multiple-interpretable improved Proximal Policy Optimization (PPO) algorithm with a few-shot technique (F-GBQ-PPO).
• Incorporated three levels of interpretability: perceptual, logical, and mathematical.
• Utilized an Apollonius circle-guided policy, a hybrid exploration policy inspired by biological motions, and quantum genetic algorithms for parameter updates.
• Implemented a few-shot technique using a multi-dimension Gaussian process to generate synthetic samples, reducing reliance on real data.

Main Results:

• The F-GBQ-PPO algorithm demonstrates increased interpretability compared to standard PPO.
• The proposed methods effectively reduce the consumption of real interaction samples.
• The few-shot technique successfully generates synthetic data, augmenting real samples to meet algorithm demands.
• Achieved improved performance in planar tracking tasks with reduced sample requirements.

Conclusions:

• F-GBQ-PPO offers a more interpretable and sample-efficient solution for planar tracking problems.
• The integration of few-shot learning significantly mitigates the need for extensive real-world data.
• The multi-level interpretability framework provides deeper insights into the tracking policy's decision-making process.