time cost is too high, the user might consider the action to be of negative utility and hence try to avoid it.
Ecology pertains to the statistical structure of the environment as expe-rienced by an individual. It involves both the immediate local task environ-ment and the environenviron-ment that the user has become familiar with over the course of a lifetime. For example, if the user knows from experience that the contact list is alphabetically ordered by last name and the person being looked for has a last name that starts with “a,” the top of the contact list is the natural place to start the search. It is important to note that ecology is considered in relation to user experience or knowledge of the environment, rather than the true structure of the environment. Although the actual structure of the environment remains the same for all users, independently of the task (unless the system is adaptive), the ecology depends on the user, not just the task.
Mechanism here refers to human capabilities. It includes cognitive and perceptual limits in the human information processing system, such as the capacity and duration of human working memory and the latencies in mo-tor movements. In our example case, parafoveal acuity and the time that it takes to fixate on an item, move the eyes, and read the name of a contact are some of the constraints. According to the AIF, the strategy space can be determined from the three components. The optimal strategy is the behavioral policy that yields the best gain in light of these: the ecology, utility, and mechanism. With the aid of the AIF and the principle of ratio-nality, we can predict the optimal strategy or the policy that a rational user would select, which usually determines the sequence of actions performed when the user interacts with a technology.
In summary, the AIF can be used to model how the user interacts with a system by specifying the utility, ecology, and mechanism factors. One can determine the strategy space and the optimal strategy by solving a machine-learning problem. In Chapter 5, I will discuss how I trained an ML model to predict the optimal strategy in an exploratory search scenario using the AIF. One of the important contributions of the AIF is that it explains why users follow different strategies to interact with the same technology.
4.2 Lookup Search as Adaptive Interaction
The adaptive nature of humans in an information search process can be ex-plained with the AIF. In this section, I explain the emergence of common lookup search strategies as adaptive interaction by considering the
seem-ingly mundane lookup search task of answering fact-related queries such as
“what is the predicted dollar-to-euro exchange rate?” Figure 4.3 illustrates this scenario as adaptive interaction.
In the latter scenario, utility is subjective and depends on what the information-seeker finds value in. It might be finding the most accurate answer or completing the task as soon as possible. If we consider a user who is involved in trading currency, accuracy of the answer might be the most important objective, but if the user is only interested in finding an approximate answer, then reasonably accurate information findable within the shortest time might be what is desired. With lookup tasks, in general, success in the task is measured in terms of task completion time [14]. Ac-cordingly, Figure 4.3 represents utility as finding a reasonable answer as quickly as possible: minimizing the task completion time.
In this scenario, the ecology element is the user’s familiarity with the statistical distribution of relevant answers on the search engine’s results page. On account of his or her experience, the user may well expect the results to be ranked in descending order of relevance to the query issued. In Figure 4.3, I illustrate the ecology distribution as a graph of user-perceived relevance of the documents plotted against document rank on the SERP.
The mechanism could be influenced by the time it takes the user to read and comprehend each item in the set of search results. The structure of the SERP too could affect the mechanism. For example, if the user can see the answer on the very first SERP, it may take less time to process the result items. On the other hand, if the user has to scroll through the SERP to find the answers, the time cost is greater. A few examples of factors that contribute to the mechanism are given in the figure.
We could predict possible strategies on the basis of AIF. There are several actions involved in a strategy, such as formulating a search query, scanning a result snippet, following a link, reading a document, judging document relevance, and terminating the search session, as indicated in Figure 4.3. The user executes these actions with different probabilities; the likelihood of a transition between any two of the actions is referred to as the transition probability. The AIF aids in finding these probabilities by solving a reinforcement learning (RL) problem. In RL, a software agent is trained to perform actions in an environment so as to maximize the reward gained from each action (Subsection 5.3.1 provides a more detailed description of reinforcement learning) [142]. Thus we can predict the strategy space and provide a logical explanation addressing why a rational user would favor one particular strategy in preference to all other possible strategies.
4.2 Lookup Search as Adaptive Interaction 47
Figure 4.3: Lookup search as adaptive interaction. The ecology shows that, operating from past experience, the user expects the results to be perfectly ranked by relevance. A hypothetical ecology function is graphi-cally represented in the figure. Utility is a function of time. Some cognitive and perceptual constraints included in the mechanism are represented in here. The strategy space is represented as a sequence of actions, where p denotes the transition probabilities for changing between actions. This rep-resentation of the strategy space is one possible approach to understanding strategy (adapted from the work of [19]).