The Positive Effect of User Faults over Agent Perception in Collaborative Settings and Its Use in Agent Design

Abstract

This paper studies the effect of user’s own task-related faults over her satisfaction with a fault-prone agent in a human-agent collaborative setting. Through a series of extensive experiments we find that user faults make the user more tolerant to agent faults, and consequently more satisfied with the collaboration, in particular compared to the case where the user is performing faultlessly. This finding can be utilized for improving the design of collaborative agents. In particular, we present a proof-of-concept for such augmented design, where the agent, whenever in charge of allocating the tasks or can pick its own tasks, deliberately leave the user with a relatively difficult task for increasing the chance for a user fault, which in turn increases user satisfaction.

Preliminary results of this work (in particular, the results of experiments T2 and T3) were presented as a poster at HAI 2021 [1].

Appendices

Appendices

A Experimental Framework Interface

Below is a screenshot of the experimental framework interface appeared to the participants (Fig. 5).

Fig. 5.

A screenshot of the “Solve the Captcha” interface.

B Experimental Treatments Comparison

Table 1. Experimental treatments.

C State Machines

Figure 6 shows the state machines associated with each treatment flow. For simplicity, arrows pointing at the end state do not appear in all relevant states. However, the game does end when the pre-specified score is reached.

Fig. 6.

State machines representing treatments flows.

D Measures

• Competence. (Question: “To what extent did you find the virtual player to be a competent partner?”) - this measure aims to capture the agent’s degree of ability, from the participant’s point-of-view.

• Satisfaction. (Question: “To what extent are you satisfied with the virtual player?”) - this measure reflects the overall user experience and degree of happiness with the agent.

• Recommendation. (Question: “To what extent would you recommend the virtual player to a friend, as a partner to work with?”) - this measure reflects the user’s loyalty.

E Complementary Graphs

While the graphs provided in the paper applied to the main measure of interest, user satisfaction, the equivalent graphs for the other two measures (user’s estimate for the agent’s competence and her willingness to recommend the agent to other users) reveal similar phenomena, qualitatively. We therefore provide these figures here.

1.1 E.1 Competence

See Figs. 7, 8, 9 and 10.

Fig. 7.

Average user competence as a function of the number of user and agent faults (equivalent to Fig. 1a in the paper).

Fig. 8.

Average user competence as a function of the number of user faults in treatment T2 (equivalent to Fig. 2 in the paper).

Fig. 9.

Average user competence as a function of the number of user faults in treatment T3 (equivalent to Fig. 3 in the paper).

Fig. 10.

Comparison of user competence in T1 and T4, providing a proof of concept for the success of designs that incorporate user’s own faults effect over her competence from the collaborative agent (equivalent to Fig. 4 in the paper).

1.2 E.2 Recommendation

See Figs. 11, 12, 13 and 14.

Fig. 11.

Average user recommendation as a function of the number of user and agent faults (equivalent to Fig. 1a in the paper).

Fig. 12.

Average user recommendation as a function of the number of user faults in treatment T2 (equivalent to Fig. 2 in the paper).

Fig. 13.

Average user recommendation as a function of the number of user faults in treatment T3 (equivalent to Fig. 3 in the paper).

Fig. 14.

Comparison of user recommendation in T1 and T4, providing a proof of concept for the success of designs that incorporate user’s own faults effect over her recommendation from the collaborative agent (equivalent to Fig. 4 in the paper).

F Participants’ Qualitative (Textual) Responses

• Agent competence per se - several subjects justified their rating solely based on the agent performance, i.e., on how they perceived their agent’s competency. Some of them expressed it explicitly (“The virtual player made several mistakes”, “He only got one wrong and was quick”), while others implicitly (“This virtual player didn’t do very well”, “My virtual partner was quick and accurate. A great team player!”).

• Agent competence vs. participant competence - many subjects explained their rating by comparing the agent competence (mostly based on its number of faults) to their own. This pattern was observed in various cases, in which the number of the agent’s faults was greater, equal, or smaller than that of the participant. For example, “The player did some extra mistake compared to me.” (two agent faults vs. one participant fault), “It was as competent as I was, only making one mistake.” (one agent fault vs. one participant fault), “Both of us solved with just one error.” (one agent fault vs. one participant fault), “played almost as good as me” (one agent fault vs. zero participant faults).

• Expectations from an automated agent - participants also correlated their ratings with their initial assumptions about and expectations from the agent. In particular, most participants emphasizing this aspect mentioned they expected the agent to be flawless. For example, “The virtual player only made one mistake, but I don’t think a virtual player should have made any mistakes at all”, “Baffled how a machine could err on such a simple task”, “If it truly was virtual, as in AI, it shouldn’t have missed any.”, “I expect a computer to be competent”. Several subjects compared the automated agent to a real person (“Virtual players is like real player”, “It kind of felt like the virtual player was a real player even though I knew it wasn’t.”), once again expressing their disappointment from agent faults.

• The complexity of the agent’s tasks - some participants correlated their rating with the relative complexity of the agent’s tasks compared to their own task. In particular, the fact that players (in T1 and T4) typically received the easier captcha seemed to had some influence on rating. For example, “He did as well I did, and it looked like he had harder puzzles”, “He only made a single mistake and had what looked like more complex captchas”, “It only made one mistake and always seemed to have longer ones to solve than I did.”.