AB
AB
AiBoss

All Categories

project

ASAL - Sakana AI, in collaboration with OpenAI and others, launched a system that uses a basic model to search for artificial life.

ASAL (Automated Search for Artificial Life) is a system that uses a basic model to automatically search for artificial life (ALife). ASAL was jointly developed by Sakana AI, MIT, OpenAI, and other institutions, and is primarily based on three...

What is ASAL?

ASAL (Automated Search for Artificial Life) is a system that automates the search for artificial life (ALife) using basic models. Developed jointly by Sakana AI, MIT, OpenAI, and other institutions, ASAL primarily utilizes three search mechanisms: supervised target search, open search, and illumination search. Supervised target search identifies simulations that produce specific phenomena or sequences of events; open search focuses on discovering simulations that can continuously generate new behaviors; and illumination search identifies simulations that exhibit diverse behaviors. ASAL has been effectively validated on several classic ALife substrates (such as Boids, particle life, life-like cellular automata, Lenia, and neuronal cellular automata), successfully discovering previously unseen life forms and expanding the boundaries of ALife research.

ASAL's main functions

  • Supervised target searchSearch for and discover artificial life simulations that can generate specific target events or event sequences, helping researchers find simulation environments that match specific phenomena.
  • Open search: To find artificial life simulations that continuously generate new behaviors and novel events, and to explore and reproduce the endless outbreaks of novelties in the real world.
  • Illumination SearchThe discovery of a group of artificial life simulations exhibiting diverse behaviors illuminates and categorizes the entire space of possible artificial life phenomena, revealing "possible forms of life".
  • Cross-matrix compatibilityIt has demonstrated broad applicability by working in a variety of different artificial life matrices, including Boids, Particle Life, Game of Life, Lenia, and Neural Cellular Automata.
  • Quantitative analysisBased on the semantic representation capabilities of the basic model, it enables quantitative analysis of artificial life phenomena that could previously only be analyzed qualitatively, providing new measurement and analysis tools.

ASAL's technical principles

  • Visual-Language Basic ModelBased on CLIP (Contrastive Language-Image Pretraining), images and natural language text are mapped to a common representation space for similarity measurement.
  • Embedding and Similarity MeasurementBased on a vision-language model, simulated images and textual cues describing the target phenomenon are embedded into the same representation space, and their similarity is measured.
  • Optimization Algorithm: Use optimization algorithms (such as genetic algorithms, CMA-ES, etc.) to adjust simulation parameters to maximize the matching degree between the representation of the target phenomenon and the simulation output.
  • Search Strategy:
    • Supervised target search: Find specific simulations based on maximizing the matching degree between the simulated generated image and the target prompt word.
    • Open search: To find open simulations that maximize the novelty of the simulated image relative to historical states in the representation space of the underlying model.
    • Illumination SearchMaximize the coverage of a set of simulations in the representation space of the underlying model, i.e., minimize the distance between each simulation and its nearest neighbor, and seek diverse simulations.

ASAL's project address

ASAL application scenarios

  • Biological evolution researchTo simulate the process of biological evolution and study how genetic variation and natural selection lead to the emergence of new species.
  • Intelligent robot behaviorTo explore the adaptive behavior and learning strategies of robots in unknown environments and improve their autonomy.
  • Game AI DevelopmentIn game development, generating non-player characters (NPCs) with complex behaviors enhances the game's immersion and challenge.
  • Ecological protection simulationSimulate the impact of different protection measures on the ecosystem to provide a scientific basis for ecological protection decisions.
  • Artistic CreationGenerate dynamic artworks that evolve based on audience interaction or changes in the external environment.
Back