Answers are grounded in this paper and cite their sources.
In recent years, reinforcement learning techniques have enjoyed considerable success in a variety of challenging domains, but are typically sample inefficient and often fail to generalize well to new environments or tasks. Humans, by contrast, are able to learn robust skills with orders of magnitude less training. One hypothesis for this discrepancy is that humans view the world in terms of objects and relations between them. Such a bias may be useful reducing sample complexity and improving interpretability and generalization. In this paper, we present a novel relational architecture which has multiple neural network sub-modules called relational units which operate on objects and output values in the unit interval. Our model transforms the input state representation into a relational representation, which is then supplied as input to a Q-learner. Experiments on a goal-seeking game with random boards show better performance over several baselines: a multiheaded attention model, a standard MLP, a pixel MLP and a symbolic RL model. We also find that the relations learned in the network are interpretable.
No takes yet. Share an insight, caveat, or question.
Kang et al. (2023) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: