Engineering Principles

The purpose of SHARPIE is to be a solution serving a wide variety of research questions across different interaction paradigms and learning modalities.

We therefore aim to follow the following principles in the design and implementation of SHARPIE

Modular

Human-AI interaction research is rapidly evolving, requiring a framework that can adapt without needing a complete overhaul.

SHARPIE aims for a strict separation of concerns by decoupling the core RL environments, the web server logic, the participant-facing web interface, and backend logging utilities.

It thereby allows researchers to swap out algorithm libraries, modify user interfaces, or plug in new interaction paradigms independently, ensuring the codebase remains clean and maintainable.

Customizable & Configurable

Every researcher, and indeed, every individual study has slightly different requirements on the interaction paradigm, the logging, and the deployment of their software.

However, there are still many across different researchers looking into the interaction of humans with (adaptive) sequential decision-making agents.

We therefore aim to build SHARPIE in a flexible way, with hooks for extending and customizing SHARPIE.

New functionalities can be adopted by SHARPIE once a reasonable mass of support for that functionality is attained. To request/suggest a new feature, please open an issue in the issue tracker to motivate the feature, where we will follow an (informal) approval process.

Multi-agent, Multi-human

Real-world human-AI interaction is rarely limited to a single agent and a single human user.

SHARPIE is built from the ground up to support multi-agent multi-human experiments, including multi-agent systems, multi-user studies, and collaborative human-AI teams composed of multiple and heterogenous agents.

The underlying communication protocols and state representations are inherently designed to scale across multiple participants and autonomous entities simultaneously.

Easily deployable

Transitioning an experiment from a local development environment to a live participant study should be seamless.

We prioritize out-of-the-box compatibility with popular cloud providers and participant recruitment platforms.

Production ready

We want the excitement to come from your experimental result, not from whether your experiment runs.

Human-subject experimentation requires software that is stable, reliable, and tested.

SHARPIE relies on established frameworks and implements industry-level quality control and testing.

Simple over complex

Given the interdisciplinary nature of human-AI research, SHARPIE must remain accessible to researchers from diverse technical backgrounds.

We favor explicit code, straightforward architectures, and minimal abstraction layers over clever or overly intricate engineering designs.

We aim to choose the path that yields the most legible code and the lowest cognitive load for the community in usage, maintenance and development.