A Closed-Loop Toolchain for Neural Network Simulations of Learning Autonomous Agents.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Neural network simulations are vital for understanding neural circuits, especially for autonomous agents learning in environments.
• Current methods using custom scripts hinder reproducibility and comparison of learning architectures.
• Closed-loop simulations, where agents interact with environments, are crucial but challenging to implement.

Purpose of the Study:

• To develop a novel toolchain integrating machine learning benchmarks with neural network simulators.
• To enable reproducible, closed-loop simulations for autonomous learning agents.
• To facilitate comparison of different reinforcement learning architectures.

Main Methods:

• Connected machine learning benchmark tools with computational neuroscience simulators (e.g., NEST).
• Developed a toolchain supporting biologically plausible neural models.
• Implemented a neuronal actor-critic architecture for reinforcement learning.

Main Results:

• Successfully trained the neuronal actor-critic architecture in standardized OpenAI Gym environments.
• Demonstrated the toolchain's functionality for closed-loop simulations.
• Facilitated performance comparison against existing reinforcement learning models.

Conclusions:

• The novel toolchain enhances reproducibility and comparability in neural network simulations for autonomous learning.
• It bridges machine learning and computational neuroscience, offering standardized environments for evaluating complex models.
• This approach supports the development and assessment of biologically plausible learning algorithms.