A Python Microframework for Training AI Agents Without Freezing the UI

A developer has built a microframework for parallel training of AI agents that tackles a classic Python problem: a frozen graphical interface during long training runs. The computation is moved into separate service processes, so the GUI stays fully responsive.

How it works

The core idea is separation: the main process handles only the GUI (wxPython), while all computation runs in separate services. While the neural network trains, the user can freely click around the interface, change settings, and watch charts update in real time. This is a standard pattern for desktop applications, but in ML training it is often ignored in favor of development speed.

The framework works with environments from Gymnasium (the standard library for reinforcement learning) and lets you add your own through plugins. Training can use a classic gradient-based approach or genetic algorithms (neuroevolution) via DEAP.

Key features

• Real-time progress visualization with Matplotlib charts
• Plugin architecture for adding new Gymnasium environments
• Neuroevolution support through DEAP genetic algorithms
• Multiprocessing paradigm for scaling across several CPU cores
• Packaging into a single .exe with PyInstaller and automated CI/CD

Practical use

The tool is useful for reinforcement learning experiments on a local machine. You can train an agent directly in the desktop application, save the results, bundle everything into a single executable, and share it with a colleague without requiring them to install dependencies. This speeds up the experimentation cycle, especially in research projects and hackathons.

What this means

The project shows that even highly specialized ML tools can be pleasant to work with if the user experience is carefully thought through. When the interface doesn't freeze, experiments move noticeably faster: there is no need to open a terminal in a separate window and run scripts from the command line.