Optimal Solution of Inductive Inference via Minimization of Unpredictable Information: Towards Artificial Superintelligence

For artificial intelligence to surpass human intelligence and reach Artificial Superintelligence (ASI), an objective and mathematical optimal solution for inductive inference, independent of subjective human evaluations or arbitrary prior probabilities, is indispensable. In this paper, we propose a novel formalization that reduces inductive inference to a deductive framework by embedding the inference target into the population as an information-less placeholder called the "Unknown" ($?$). In this method, the probability distribution and the unknown rate of the target are derived based on the spatial relationships between known samples and the target. Furthermore, to mathematically eliminate the combinatorial explosion (overfitting) associated with infinite variable generation and coordinate transformations, we redefine the optimal solution of inductive inference as the minimization of the "total unpredictable information," which is the sum of the "information content of the rule" and the "entropy of the inference result." By calculating the rule's information content as the number of symbol selections in the sum-of-products standard form, and by defining entropy through a parameter-free general formula based on the total number of samples, we completely eliminate dependence on programming language syntax and physical resolution. This theory resolves the trade-offs among inference purity, sample size, and rule complexity without any arbitrariness. Implementing this objective evaluation criterion as the absolute objective function for the AI's "System 2" (rational reasoning) paves a definitive path toward realizing ASI that transcends human cognitive biases and autonomously discovers true laws.