Modeling of autonomous problem solving process by dynamic construction of task models in multiple tasks environment

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Traditional reinforcement learning (RL) agents struggle to adapt to new, similar tasks, requiring relearning from scratch.
• This inefficiency in quick action learning hinders effective decision-making in dynamic, real-world environments.

Purpose of the Study:

• To propose a novel action learning technique for reinforcement learning agents.
• To enable agents to efficiently solve a set of similar tasks in a multi-task environment by leveraging previously acquired knowledge.

Main Methods:

• A model-based reinforcement learning approach is utilized.
• A task model is constructed by integrating primitive local predictors to forecast environmental and task dynamics.

Main Results:

• The proposed method demonstrates the ability to quickly solve similar tasks by reusing learned knowledge.
• Computer simulations using a varied ping-pong game validated the effectiveness of the technique.

Conclusions:

• The developed technique significantly enhances the agent's capability for rapid adaptation to related tasks.
• This approach addresses the limitations of traditional RL in scenarios requiring quick action learning and decision-making.