PAC Learning of Causal Trees with Latent Variables
Authors: Prasad Tadepalli, Stuart J. Russell9774-9781
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental Results Unfortunately the worst-case PAC bounds tend to be very pessimistic in general and are not useful to predict the learning curves. To examine the empirical performance of our algorithm, we evaluated it on datasets generated from synthetic target causal trees. |
| Researcher Affiliation | Academia | 1 School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, Oregon, 97331 2 Division of Computer Science, University of California, Berkeley, California, 94720 |
| Pseudocode | Yes | Algorithm 1 Normalizing a Causal Tree, Algorithm 2 The Bottom-up Tree-Building Algorithm, Algorithm 3 Filling in the CPT of a Node |
| Open Source Code | No | The paper does not provide an explicit statement or a link to open-source code for the described methodology. |
| Open Datasets | No | To examine the empirical performance of our algorithm, we evaluated it on datasets generated from synthetic target causal trees. Each is a full binary tree with depth 4 and 16 leaf nodes. At each node we randomly selected a CPT, where each probability in the table is chosen independently from a uniform distribution in the union of the two intervals (0.0, 0.1) and (0.9, 1.0). The paper does not provide a specific link, DOI, repository name, or formal citation for accessing these datasets. |
| Dataset Splits | No | The paper mentions 'training' and 'test' examples but does not explicitly describe a validation dataset split or a methodology like cross-validation for hyperparameter tuning. |
| Hardware Specification | No | The paper does not provide specific hardware details (like GPU/CPU models, memory, or cluster specifications) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers. |
| Experiment Setup | Yes | Each is a full binary tree with depth 4 and 16 leaf nodes. At each node we randomly selected a CPT, where each probability in the table is chosen independently from a uniform distribution in the union of the two intervals (0.0, 0.1) and (0.9, 1.0). The input example distribution P is uniform. During the training of each batch, the algorithm also asks experimental queries, which are answered by referring to the target tree. When a query is answered, its response is stored so that it is never asked again. After each batch of training the learned tree was evaluated on 40 test examples which were chosen randomly from the pool of examples not queried during training. We report in Table 1 the average results of 30 different training sets on 10 target trees, one per each row. The first two columns are based on an algorithm that knows the tree structure. It is identical to Algorithm 2 with two differences: (1) the candidate list of trees constructed in line 4 will only include the correct subtrees at the next level, and (2) the test if f(T, S) < θ in line 13 is removed and treated as true since we are only considering correct subtrees. The last 3 columns are based on our algorithm. Columns 1 and 3 report the number of average unique queries asked on each target tree in the two scenarios. To achieve comparable test error rates, the algorithm was run with 5 random examples when the tree is not given and 1 random example when the tree is given. |