Summary Of The Paper:
The paper proposed an algorithm to model trust numerically in parallel with the policy learning process in multi-agent scenario. The algorithm resembles an overlay on any reinforcement learning algorithm with modification using a dynamic trust reward, depending on trust levels. The proposal is inspired by dopaminergic system and reaffirms the Dunning-Kruger effect (DKE) in societal studies.
Main Review:
Pros: 1. The narrative perspective is interesting and the analogy from trustor-trustee advisory is intuitive, and easy to relate to the technical formulations. 2. More contemporary concept like distributional reinforcement learning is employed for learning multiple trust levels, and the authors proposed a novel dynamic trust reward based on the change in trust level during the learning process. 3. This paper provides comprehensive experiments on evaluating hyperparameter sensitivity as well as simulating the DKE overconfidence effect and reproduce the results from other works, adding confidence to the experiment findings.
Cons: 1. Although it is clear that the goal is to construct a trust model in multi-agent scenario, but the ultimate motivation in the bigger picture perspective is not addressed clearly. For example, what is the benefit or application of constructing a trust model? 2. From a policy performance standpoint, there is no comparison between the performance when the trust model is present/absent (ablation study). Understand that this may not be the primary goal of this paper (also somehow related to the first question), but this would add to the completeness and perhaps spark future interests among readers.
Questions for clarification: 1. In the background and introduction section, the definitions and terminologies of trust, confidence and available actions are explained as an integral part of the social systems structure. What role does the “confidence” play in the proposed modelling of trust-decision making process? It seems the proposed method is based primarily on “trust” and “confidence” is an 2. In section 1.3, “… shared narratives tend to become more liquid. As a consequence, confidence is decreasing to the benefit of trust.” This sounds interesting, but is not exactly clear. Can you elaborate more on this? High self-trust level seems to imply higher confidence in your discussion, but why does this statement observe an inverse relationship between the two? 3. Is there an explanation or hypothesis on why overconfident is more pronounce in higher complexity problems (especially when the paper has a heavy psychological-inspired tone)? I feel it is better to include at least a line of hypothetical explanation rather than leaving it completely as an open question.
Some typos: 1. Section 1.2 paragraph 3 last sentence: condidence -> confidence 2. Section 3.1 paragraph 2 second sentence: [-1, 1[ -> [-1, 1]
Summary Of The Review:
The paper takes on an interesting perspective in trust modelling while connecting the biological, humanities, psychological and societal studies inspirations to the proposed method. The technical formulations are easy to follow and sound.
Correctness: 3: Some of the paper’s claims have minor issues. A few statements are not well-supported, or require small changes to be made correct.
Technical Novelty And Significance: 3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.
Empirical Novelty And Significance: 3: The contributions are significant and somewhat new. Aspects of the contributions exist in prior work.
Flag For Ethics Review: NO.
Recommendation: 6: marginally above the acceptance threshold
Confidence: 3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.