Table 2 3.2 Learning of the causal Bayesian networks
Related Figures (8)
Fig. 1. Race-like environment considered in this study for our self-driving car (in bright red). Fig. 2. Predefined locations for other vehicles around the self-driving car (dashed red lines indicate the space the self-driving car occupy on each lane). When the vehicle is on the left (resp. right) lane, only the locations Northwest, Northeast, East and Southeast (resp. Northeast, Northwest, West and Southwest) are meaningful. Table 1. Number of actions in state-action pairs labeled as potential collisions and non-potential collisions in the integrated dataset. Table 2. Number of unique state-action pairs labeled as non-potential collisions and potential collisions (“safe” and “unsafe”, respectively). Table 3. Number of unique state-action pairs as a function of the action used fo training and testing each cBN. are used for training and testing) does not represent a major concern in this setup. This is because the core of the evaluation focuses on comparing condi- tional probabilities under interventions for a counterfactual model, rather than studying the generalization capabilities of PLTNs to previously unknown data. Despite this, we consider constructing cBNs using training sets of varying sizes for comparison purposes. To achieve this, we randomly selected 1%, 50%, and 100% of the training dataset for structural and parameter learning of the cBNs. Table 3 summarizes the number of unique state-action pairs used for training and testing for each driving action and cBN. Fig. 3. Causal Bayesian network (cBN) proposed in this work. Table 4. Number of groups with a unique minimum probability value and the numbe! of ties for first place (ranging from 2 to 5) across all groups and counterfactual models 4 Evaluation and results Table 5. Number of actions selected as the optimal intervention in the three counter- factual models. Table 6. State-action pairs with 5 tied actions. The first two pairs, from top to bottom, correspond to the counterfactual model trained with 1% of the data, and the last one is the same for the models trained with 50% and 100% of the data (all actions other than the observed are safe). 4.1 Discussion
