What Planning Problems Can A Relational Neural Network Solve?
Authors: Jiayuan Mao, Tomás Lozano-Pérez, Josh Tenenbaum, Leslie Kaelbling
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We consider two families of problems: one predicted to require finite depth and one predicted to require unbounded depth. For all tasks, we use Neural Logic Machines [Dong et al., 2019] as the model; we set the number of hidden neurons in the MLP layers to be sufficiently large (64 in all experiments) so that it is not the bottleneck for network expressiveness. We show the average performance across 3 random seeds, with standard errors. Assembly3: finite depth circuits. The domain Assembly3 contains n objects. ... Shown in Table 2, we train models on environments with 10 objects and test them on environments with 10, 30, and 50 objects. The first model is not able to learn a policy for the given task, while the second exhibits perfect generalization. Logistics: unbounded depth circuits. ... Shown in Table 2, the first model fails to generalize to larger graphs due to its limited circuit depth. By contrast, the second model with adaptive depths generalizes perfectly. Blocks World-Clear: unbounded depth circuits. ... Shown in Table 2, although both policy networks accomplish the task with a 1.0 success rate (because there is a trivial policy that randomly drops clear blocks onto the table), the policy whose depth depends on the number of blocks successfully finds a plan that is significantly shorter than fixed-depth policy. |
| Researcher Affiliation | Academia | Jiayuan Mao1 Tomás Lozano-Pérez1 Joshua B. Tenenbaum1,2,3 Leslie Pack Kaelbling1 1 MIT Computer Science & Artificial Intelligence Laboratory 2 MIT Department of Brain and Cognitive Sciences 3 Center for Brains, Minds and Machines |
| Pseudocode | Yes | Algorithm 1 Plain backward search. function bwd(s0, A, goal_set) ... Algorithm 2 Serialized goal regression search. function s-grs(s0, R, g, cons) ... Algorithm 3 Serialized goal regression search with multiple optimal path tracking. function grsopt(s0, g, cons) |
| Open Source Code | Yes | Code. We release the code for reproducing all our experimental results in https://github.com/ concepts-ai/goal-regression-width. |
| Open Datasets | No | For the Assembly3 task, we first uniformly sample the set of type-A, type-B, and type-C object. For the Blocks World-Clear task, we use the same random sampler for initial configurations as Neural Logic Machines [Dong et al., 2019]. For the Logistics task, we first randomly generate a tree rooted at the start node. The paper describes custom-generated environments and sampling methods rather than using or providing access to a publicly available dataset with concrete links or citations. |
| Dataset Splits | No | We train models on environments with 10 objects and test them on environments with 10, 30, and 50 objects. The paper describes training on smaller-sized environments and testing on various sizes for generalization, but does not specify explicit training, validation, and test dataset splits with percentages or counts for a fixed dataset. |
| Hardware Specification | Yes | Compute. We trained with 1 NVIDIA Titan RTX per experiment for all datasets, from an internal cluster. |
| Software Dependencies | No | The paper mentions using Neural Logic Machines as the model and the Adam optimizer, but does not provide specific version numbers for any software dependencies like Python, PyTorch, or other libraries. |
| Experiment Setup | Yes | Additional hyperparameters. The breadth, width, and hidden dimension parameters for neural networks have been specified in the experiment section and they vary across tasks. For all experiments, we use the Adam optimizer to train the neural network, with a learning rate of ϵ = 0.001. Additionally, we set β1 = 0.9 and β2 = 0.999. |