An international team has developed an artificial intelligence model capable of planning actions without systematically exploring every possible solution. Inspired by mechanisms associated with the hippocampus, the system combines cognitive maps, random computations, and reusable elements to progress toward a goal.
Cognitive maps represent relationships between objects or situations in a geometric form. They provide the model with a sort of general direction. It can then test intermediate steps, keep those that bring it closer to the goal, and quickly abandon less promising avenues.
This approach also relies on neural sampling. The system generates several possible scenarios without calculating each trajectory to its end. It then evaluates their orientation in the cognitive map before pursuing the options deemed most useful.
The researchers add a so-called compositional organization. Information and plans are divided into elements that can be reused in different situations. This principle allows the model to assemble a new strategy from already available components, rather than completely restarting its learning.
The system was tested on three categories of tasks. It had to navigate a two-dimensional space, orient itself in an abstract environment with multiple dimensions, and then assemble and disassemble a shape made of different parts.
According to the authors, the model can also adapt when the situation changes, without requiring full retraining. This capability could be of interest to robots, autonomous vehicles, or devices operating locally with limited resources.
The researchers also present their method as far less energy-intensive than multilayer neural networks or large language models.
The team is therefore not seeking to reproduce all the functions of a general-purpose artificial intelligence. Instead, it proposes a specialized architecture for planning and problem-solving.