1fdc0ee9d95c71d73df82ac8f0721459-AuthorFeedback.pdf

Wewould liketothank the reviewers for their comments, and inparticular appreciate the insightful suggestions for1 futureworkbyR1andR2. Table 1 shows that GFSA has excellent sample efficiency, and Fig 1 shows an9 exampleofmany-stepbehavior. Incontrast, RL approaches produce stochastic discrete samples. We agree with R1 that applying GFSA ideas to other domains is an exciting next step. Conditioning on the initial node is straightforward because we can still marginalize out the agent'shistory50 whilesolving forthedistribution.