Artificial intelligence system? general before learning new environment. In contrast, creativity is an innate human trait: they have the ability to map their previous experiences to novel situations.
Inspired by this human ability, a recent article on arXiv.org introduces a theoretical model of Deep Reinforcement Learning that aims to allow artificial agents to learn heterogeneous, possible concepts. generalization and relational communication.
The new model constructs a hierarchical conceptual space consisting of objects, affordability, and representation of their relationships. When the concepts retrieved do not yield satisfactory results, the concepts can be recombined to form new and validated concept representations in the same reinforcement learning setting. The above relationship level is encoded with the reward into abstract conceptual states. This information provides the agent with multiple perspectives on a concept, allowing for the transfer of learning between tasks and goals.
This paper introduces a computational model of creative problem solving in deep reinforcement learning agents, inspired by cognitive theories of creativity. The AIGenC model is intended to allow artificial agents to learn, use, and create transferable representations. AIGenC is embedded in a deep learning architecture consisting of three main components: concept processing, reflective reasoning, and concept matching. The first component extracts objects and affordability from sensory input and encodes them in a conceptual space, represented as a hierarchical graph structure. Concept representations are stored in a dual memory system. The temporal and target orientation information acquired by the agent during deep reinforcement learning enriches the representations creating a higher level of abstraction in the conceptual space. In parallel, a process similar to reflective reasoning detects and recovers from task-related memory concepts by a matching procedure that computes the similarity between the current state and the structure memory graph. After completing an interaction, rewards and information are temporarily added to the graph structure, creating a higher level of abstraction. If reflective inference does not provide a suitable solution, an associative process is introduced to generate new concepts by combining past information. We discuss the model’s ability to yield better out-of-distribution generalization in artificial agents, thus moving towards general artificial intelligence. To the best of our knowledge, this is the first computational model that, beyond mere formal theories, offers a solution for creative problem solving in a deep learning architecture.
Research articles: Catarau-Cotutiu, C., Mondragon, E. and Alonso, E., “AIGenC: Generative AI Through Innovation”, 2022. Link: https://arxiv.org/abs/2205.09738
