We present a comparative study of probabilistic logic factored Markov decision processes (PL-fMDPs), classification and regression trees (CART), and multilayer perceptrons (MLPs) for behavior selection in self-driving cars. While CART and MLPs are widely used in decision-making, PL-fMDPs have been recently proposed for autonomous behavior selection with promising results. We carried out three main tests to evaluate these models: (i) learning and testing with examples taken from a simulated self-driving vehicle in a race-like scenario, (ii) comparison with actions of human drivers, and (iii) navigation of a self-driving car in two adverse and unknown road scenarios. In the first and third tests, CART slightly outperformed MLPs, and both narrowly surpassed PL-fMDPs. However, PL-fMDPs showed noticeably better alignment with the decisions of human drivers in the second test, at the cost of increased learning and testing time. The results demonstrate the competitiveness of all three approaches, although the definitive superiority of any one model was not observed across all evaluation contexts. This reinforces the need to explore hybrid approaches that combine symbolic, probabilistic, and connectionist models to leverage their respective strengths and mitigate their limitations for autonomous driving in self-driving cars.

We investigate how causal DAG learning algorithms and structural assumptions influence counterfactual decision safety. Four structure learning regimes are compared: expert-guided edge-constrained HC+BIC, unconstrained HC+BIC, MMPC+HC+BIC, and the PC-Stable algorithm. Evaluation is conducted using a leave-one-state-out protocol over a fully enumerated state–action space in a controlled offline autonomous driving setting. The environment is characterized by seven Boolean state variables and six actions, allowing us to disentangle the effects of learning strategies on counterfactual decisions. All models are implemented as probabilistic logic twin networks (PLTNs), with additional sensitivity analysis across parameter configurations. The learning regimes produce markedly different counterfactual decisions. Edge-constrained HC+BIC recommends a diverse set of safe actions, while unconstrained HC+BIC yields fewer but consistently safe alternatives. MMPC+HC+BIC frequently fails to identify safe actions, often associated with weak connectivity of the outcome variable. PC-Stable produces varied recommendations but may include unsafe actions, which is linked to incorrect edge orientations between actions and outcomes. These findings show that structure learning choices and prior knowledge influence counterfactual decisions through the learned structure, affecting the identification of safe alternatives in safety-critical applications.