Abstract
This paper presents a novel neural network framework for solving complex abductive reasoning problems. By integrating Potts spin models with Mean Field Theory (MFT) and an annealing process, the proposed system efficiently identifies the most parsimonious set of causes for observed symptoms. The approach effectively manages logical constraints such as incompatibility and coverage, outperforming traditional models in both speed and accuracy on large-scale diagnostic tasks.
1. Introduction
Causal reasoning, or abduction, is the process of inferring the best explanation for a set of observations. In domains like medical diagnosis, this task is often NP-complete due to the exponential growth of potential cause combinations and the presence of conflicting evidence. Traditional neural networks often struggle with local minima, while stochastic methods like Simulated Annealing are computationally expensive. This work addresses these challenges by mapping logical abduction into an energy-minimization problem solvable via Mean Field Annealing.
2. Methodology
The core of the proposed system is an energy function that encodes the requirements of a valid diagnostic scenario. It rewards the coverage of observed effects and penalizes the inclusion of incompatible causes or causes that imply absent effects. The model utilizes Potts spins, which naturally implement a "winner-takes-others" competition, ensuring that every effect is explained by at least one cause.
The system employs Mean Field Annealing (MFA), replacing discrete binary states with continuous thermal averages. By gradually lowering the "temperature," the network avoids local traps and converges toward a global optimum. This process allows the network to handle complex interactions, such as those defined in the medical domain ruleset shown in Figure 1.
3. Experimental Evaluation
The model was evaluated on the Causal Problem Medical Domain (CPMD) and a large-scale synthetic dataset. In the CPMD example, the network successfully navigated conflicting symptoms. For instance, while Pneumonia could explain several symptoms, its incompatibility with other observations led the network to favor a combination of Laryngitis and Sinusitis.
As shown in the activation plot (Figure 2), the network's internal states "crystallize" into a definitive solution as the time-counter (representing the cooling process) progresses. In a large-scale test of 1000 cases involving 26 diseases and 56 symptoms, the Potts MFT model demonstrated superior efficiency and robustness compared to existing abductive models, finding more optimal scenarios in significantly less CPU time.
4. Conclusion
The Potts Spin MFT network provides a powerful bridge between logical reasoning and statistical mechanics. By treating diagnostic constraints as physical potentials, it achieves high-performance abduction. Future work will focus on extending the model to handle complex interactions like causal cancellation and improving the empirical estimation of critical temperatures.
Summary updated