Tuesday 08 April 2025
The quest for intelligent machines that can think and act like humans has been a longstanding challenge in artificial intelligence research. Recently, a team of researchers has made significant progress towards achieving this goal by developing a hierarchical neuro-symbolic decision transformer.
This innovative system combines the strengths of symbolic planning with the flexibility of neural networks to create a robust and efficient framework for decision-making. The idea is to use a high-level symbolic planner to generate a sequence of discrete sub-operators, which are then refined by a lower-level control logic implemented using a transformer architecture.
The researchers employed this approach in two case studies: a single-key door environment and a multi-goal key-door environment. In the first scenario, the agent must navigate through a 5×5 grid to reach a goal cell while picking up a single key and opening a locked door. The system successfully demonstrated improved success rates and efficiency over a purely neural approach.
In the second case study, the researchers introduced multiple keys, doors, and collectible items that must be acquired before reaching an exit cell. This environment presents a more complex challenge for decision-making, but the hierarchical neuro-symbolic framework again showed promising results, outperforming a pure neural approach in success rates and efficiency.
The key innovation of this system lies in its ability to combine symbolic planning with neural networks. The high-level planner generates a sequence of discrete sub-operators that are then refined by the lower-level control logic implemented using a transformer architecture. This allows the system to leverage the strengths of both approaches, achieving robustness and efficiency in decision-making.
The researchers used a standard decoder-only transformer stack, consisting of multiple layers with multi-head self-attention, layer normalization, and feed-forward blocks. The sub-goal token is conditioned on the recent states and actions, allowing the model to differentiate between different sub-goals. The system is trained using offline trajectories collected via random exploration.
The experiments demonstrated that the hierarchical neuro-symbolic decision transformer can effectively solve complex decision-making tasks by combining symbolic planning with neural networks. This approach has significant potential for real-world applications, such as autonomous vehicles, robots, and other systems that require intelligent decision-making.
Overall, this research represents a major step forward in developing intelligent machines that can think and act like humans. The hierarchical neuro-symbolic decision transformer demonstrates the power of combining symbolic planning with neural networks to achieve robust and efficient decision-making.
Cite this article: “Revolutionizing Artificial Intelligence: Hierarchical Neuro-Symbolic Decision Transformers for Efficient and Explainable Control”, The Science Archive, 2025.
Artificial Intelligence, Neural Networks, Symbolic Planning, Decision Transformer, Hierarchical Framework, Multi-Goal Environments, Autonomous Vehicles, Robots, Intelligent Machines, Robotics
Reference: Ali Baheri, Cecilia O. Alm, “Hierarchical Neuro-Symbolic Decision Transformer” (2025).
