Contrastive Behavioral Similarity Embeddings for Generalization in Reinforcement Learning

Authors: Rishabh Agarwal, Marlos C. Machado, Pablo Samuel Castro, Marc G Bellemare

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that PSEs improve generalization on diverse benchmarks, including LQR with spurious correlations, a jumping task from pixels, and Distracting DM Control Suite.
Researcher Affiliation Collaboration Rishabh Agarwal Marlos C. Machado Pablo Samuel Castro Marc G. Bellemare Google Research, Brain Team {rishabhagarwal, marlosm, psc, bellemare}@google.com Also at Mila, Université de Montréal. Now at Deep Mind.
Pseudocode Yes Algorithm 1 Contrastive Metric Embeddings (CMEs) and J. PSEUDO CODE, including functions like def metric_fixed_point and def contrastive_loss.
Open Source Code No We use the open-source code released by Sonar et al. (2020) for our experiments.
Open Datasets Yes Jumping task from pixels (Tachet des Combes et al., 2018), LQR with spurious correlations (Song et al., 2019), and Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b).
Dataset Splits Yes We split the problem into 18 seen (training) and 268 unseen (test) tasks... For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the wide grid (Figure 2a) and pick the parameters with the best validation performance.
Hardware Specification No The paper does not specify the exact hardware (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No All agents are built on top of SAC (Haarnoja et al., 2018) combined with Dr Q (Kostrikov et al., 2020)...
Experiment Setup Yes Table G.2: Common hyperparameters across all methods for all jumping task experiments. Table G.3: Optimal hyperparameters for reporting results in Table 1. Table G.4: Optimal hyperparameters for reporting results in Figure 5.3. Table G.5: Optimal hyperparameters for reporting ablation results in Table 2.