Clarification on ε = 1 in final best-belief agent selection

[zzp-elio]

Hi, thank you for releasing the HGM paper and code.

I have a clarification question about the final agent selection strategy. In Section 3.3, the paper defines the best-belief agent using the epsilon percentile of the utility posterior, and later states:

In all experiments, we employ HGM with an exploration-exploitation scheduler B / b, where b is the remaining budget, epsilon = 1, and alpha = 0.6.

I am a bit confused about the interpretation of epsilon = 1.

If epsilon = 1 is interpreted as the 100th percentile of a Beta posterior, then the percentile value would be 1 for all agents, so it would not distinguish between different agents.

Did you intend epsilon = 1 to mean the 1st percentile of the posterior, i.e. epsilon = 0.01 in probability notation?

Thanks for the clarification!

[Kenan-jeff-Jiang]

I have the same question. If ε = 1 is interpreted as the 100th percentile of the Beta posterior, then every agent’s posterior quantile would be 1, since the Beta distribution is supported on [0, 1]. In that case, the final best-belief criterion would not distinguish between agents.

Could you clarify whether ε = 1 was intended literally, or whether the implementation uses a different convention/value for final selection?