Learning sparse relational transition models

Authors: Victoria Xia, Zi Wang, Kelsey Allen, Tom Silver, Leslie Pack Kaelbling

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 EXPERIMENTS We apply our approach, SPARE, to a challenging problem of predicting pushing stacks of blocks on a cluttered table top. We describe our domain, the baseline that we compare to and report our results. Figure 4: (a) In a simple 3-block pushing problem instance, data likelihood and learned default standard deviation both improve as more deictic references are added. (b) Comparing performance as a function of number of distractors with a fixed amount of training data. (c) Comparing sample efficiency of SPARE to the baselines. Shaded regions represent 95% confidence interval.
Researcher Affiliation Academia Victoria Xia Zi Wang Kelsey Allen Tom Silver Leslie Pack Kaelbling Equal contribution. Massachusetts Institute of Technology, 77 Massachusetts Ave., Cambridge, MA 02139. {vxia,ziw,krallen,tslvr,lpk}@mit.edu.
Pseudocode Yes Algorithm 1 Greedy procedure for constructing Γ.
Open Source Code No The paper does not include any explicit statement about releasing the source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets No We simulate this 3D domain using the physically realistic Py Bullet (Coumans & Bai, 2016 2018) simulator. The paper describes generating data by sampling problem instances and action parameters, rather than using a pre-existing publicly available dataset with concrete access information.
Dataset Splits Yes We held out 20% of the training data as the validation set.
Hardware Specification No The paper does not provide specific hardware details (like GPU/CPU models, memory, or specific computing environments) used for running the experiments.
Software Dependencies No The paper mentions software like 'Keras' and 'Adam optimizer' without specific version numbers. It also refers to 'PyBullet (Coumans & Bai, 2016 2018)', which is a range and not a single, specific version number for reproducibility.
Experiment Setup Yes Predictors for the templates approach were trained for 1000 epochs each with a decaying learning rate starting at 1e-2 and decreasing by a factor of 0.6 every 100 epochs. The GNN was trained using a decaying learning rate starting at 1e-2, and decreasing by a factor of 0.5 every 100 epochs. A total of 900 epochs were used.