We're pleased to introduce Automated Discovery Tool, a software for assisted and automated discovery of patterns in the exploration of complex systems.
Automated Discovery Tool is a software package developed in the Inria FLOWERS research team which provides an integrated solution for studying complex systems through curiosity-search methods, consisting of a user-friendly Web UI and an extensible Python library for user-defined experimentation systems and search algorithms.
Searching the configuration space of complex systems is often done manually, i.e., by a human who individually identifies interesting patterns or behaviors of the system. Automated Discovery Tool thus assists in automating this exploratory phase of researching a new system which is theorized to be capable of interesting, yet unknown behavior. This is the case for many projects in the natural sciences and elsewhere. For example, physicists and chemists may use the tool study the emergence of novel structures and materials from a physical system, or digital artists and designers may use the tool to automatically generate or iterate on existing designs during the creative process.
Please note that this software is currently in an alpha stage of development: it is functional and has been used internally at Inria FLOWERS to study cellular automata since 2021, but may not have features which are convenient for different workflows. For more details on the development of Automated Discovery Tool, see the following usage and technical section.
In the above demo, the tool is used to discover life-like propagating patterns in a cellular automata simulation, showcasing how a researcher can specify a series of experiments and monitor their results through both the Web UI and an integrated Jupyter notebook.
The software was designed and maintained with contributions from Chris Reinke, Clément Romac, Matthieu Perie, Mayalen Etcheverry, Jesse Lin, and other collaborators in the FLOWERS team.
The high-dimensional phase space of a complex system poses many challenges to
study. In particular, it is often desirable to explore the behavior space of
such systems for interesting behaviors without knowing a priori the precise
quantities to look for. As such, a class of algorithms based on intrinsic
motivation or "curiosity" has been proposed in
Reinke et al., 2020 and extended in
e.g., Etcheverry et al., 2020
Such curiosity algorithms enable a system to automatically generate a
learning curriculum from which it learns to explore its behavior space
autonomously in search of interesting behaviors, originally proposed in the
context of robotic agents learning to interact with their environment
in an unsupervised manner, as in
Oudeyer et al., 2007.
In practice, dealing with such ill-posed and/or subjective search tasks requires significant human oversight. For this reason, our Automated Discovery Tool proposes a software package for both :
- the implementation of such experimental pipelines for arbitrary systems and search methods, and
- the human-supervised exploration of such systems.
The repo comes with an existing implementation of the Lenia system which can be explored using the curiosity search algorithms described. The Python API described in the technical documentation can be used to add custom systems and (optionally) search algorithms.
The application uses Docker and will install
on first run of the start_app.sh script. Docker enables cross-platform
compatibility, and the application has been tested on MacOS and Linux.
Please see the online documentation at https://developmentalsystems.org/adtool/.
Please note that as the software is currently in an alpha stage as of January 2023, breaking changes to the API may occur. This is due to the current progress towards two key developmental milestones:
- Streamlining of the API, allowing checkpointing and restoring of experiments with arbitrary data, à la Git
- Implementation of a human interface for interaction with the search process itself, allowing custom human intervention at each step taken
Due to the software currently being in an early stage of development, there may be questions which are not immediately answered by documentation, and you may instead direct them to Jesse Lin who is happy to respond.