Observation & Confirmation of Causality for Active symbolic Modeling with Information theory (OCCAMI)

Symbolic models are generally preferable to non-symbolic models (e.g. neural networks with billions of weights, which are essentially lookup tables), as they are more compact and parsimonious, so they: require less storage to remember them require less CPU power to execute them (post-training) require less communications bandwidth to share them with other friendly agents extrapolate & generalize better to “black swans” absent from training data sets (Ockham’s Razor) are much more explainable Symbolic causal models are even better: humans are not statistical, but causal, thinkers, so better understand causal explanations causal models thus can be analyzed [by experts] on physical principles, justifying trust (or rejection) they guide RL agents to opportunities to intervene in the world to achieve a desired effect However, the search space of symbolic models is theoretically unbounded. Genetic Programming (GP) can search this space, but at theoretically unbounded, and practically enormous, computational cost. To make the problem tractable, it can be addressed in several self-supervised Active Learning steps: detect & estimate the strengths of arbitrary [multivariate] [nonlinear] statistical associations among observed variables (this can be accomplished using information theory as in the PI’s prior ELITE2 work) among these, estimate Granger causality (again using an information theoretic formulation) from those with strong Granger causality, filter out those that fail automatable checks for plausible causality automatically apply Pearl’s back- & front-door criteria & adjustments to colliders, confounders &mediators confirm observed plausibly causal direct links with either human expert analysis experimental (Reinforcement Learning) interventions construct a causal graph from the confirmed direct causal links search for the forms of the causal dependencies only on the relevant direct causal links optimize the numerical parameters in the discovered symbolic formulae Thus, before searching for the parameters of a relationship, determine the symbolic form of that relationship, and before that determine whether a causal direct relationship even exists. This reduces each stage of the search to a space of much lower dimensionality. Determining the integer numbers of real state variables involved, the minimally sufficient length of a lag or successive differential vector, etc. of a passively observed physical system, we call Reasoning About  Z Over R (RAZOR). Together with the subsequent model learning steps listed above, we propose this as “Observation & Confirmation of Causality for Active symbolic Modeling with Information theory (OCCAMI)”. Critical Technologies Inc. & Evolved Analytics LLC will implement and commercialize OCCAMI  this by extending the latter’s well-established software tool DataModeler and will support military transition by “agentifying” OCCAMI as headless code embeddable in autonomous cyber-physical systems inc. drones.