A mathematical model for the two-learners problem.

Area of Science:

• Machine Learning
• Human-Computer Interaction
• Computational Neuroscience

Background:

• Co-adaptive learning involves systems where multiple agents learn and adapt in response to each other.
• Understanding the dynamics of human-machine interaction is crucial for developing effective collaborative systems.
• Previous research in Brain-Computer Interfaces (BCI) has explored human-machine adaptation, but a generic theoretical formulation was lacking.

Purpose of the Study:

• To present the first generic theoretical formulation of the co-adaptive learning problem.
• To investigate the learning dynamics of interacting human and machine agents.
• To identify optimal learning parameters for efficient human-machine collaboration.

Main Methods:

• Developed a simple linear model of co-adaptive learning with a joint loss function for human and machine agents.
• Derived theoretical learning rules for both human and machine agents to enable computational simulations.
• Validated the theoretical model through a real-world experimental study involving human-machine cursor control under distortion.

Main Results:

• Simulations revealed a rich, complex structure in the co-adaptive learning ecosystem.
• Identified a 'sweet spot' of mid-range learning rates for the machine, leading to rapid convergence, largely independent of human learning rates.
• Experimental results confirmed that mid-range machine learning rates yield the best performance and user experience in a practical cursor control task.

Conclusions:

• The theoretical framework provides a basis for understanding and simulating co-adaptive learning systems.
• Mid-range machine learning rates are critical for efficient and stable human-machine collaboration.
• Findings align with and extend previous observations in the Brain-Computer Interface literature.