On the Effectiveness of Strategies
in Different Types of Persuasion Dialogue Settings
There are many situations in which agents may need to persuade other agents. This can be formalised as persuasion dialogues, in which agents exchange arguments represented by abstract entities in order to establish whether a specific goal argument should be deemed acceptable based on an underlying logical framework. An active area of research is how to find strategies for persuasion dialogues that prescribe which arguments a proponent should assert in which order to cause an opponent to accept a given goal argument. The proponent may have uncertain knowledge about what arguments are known to the opponent in the form of an opponent model containing different possible agent models, and may or may not have any knowledge of the opponent’s behaviour during the dialogue.
A strategy is successful if using it results in the goal argument being acceptable. If the proponent has knowledge about the behaviour of the opponent, then it is possible to calculate the probability of success of a strategy. However, if no knowledge about the behaviour of the opponent is available, then the probability of success cannot be calculated and the performance of a strategy has to be measured differently. In these cases, the probability of guaranteed success of a strategy, i.e. the probability that a dialogue will terminate successfully when this strategy is used regardless of how the opponent behaves, can be used to evaluate the performance of strategies. We discuss the differences between probability of guaranteed success and probability of success, and argue that probability of guaranteed success is a valuable measure for evaluating the performance of strategies in persuasion dialogues due to its independence from assumptions or knowledge about the behaviour of the opponent. We also introduce an upper bound on probability of guaranteed success, which allows for more meaningful analysis of the performance of strategies with regards to probability of guaranteed success.
Finding simple strategies, which consist of asserting sets of arguments in a predefined order independent of the opponent’s behaviour, can be an effective way to optimise for probability of guaranteed success while significantly reducing the search space of all available strategies. We investigate the performance of simple strategies in different types of persuasion settings in order to analyse under which circumstances optimal simple strategies are optimal general strategies in terms of guaranteed success. In order to better understand the performance of simple strategies in different types of persuasion settings, we introduce the notion of conflict in persuasion settings, which makes it possible to group settings into three different categories that align with the structure of optimal simple strategies: no conflict, simple conflict, and complex conflict. We show that in settings with no conflict or with simple conflict, optimal simple strategies cannot be outperformed when considering the probability of guaranteed success. Thus, in these settings it is sufficient to only consider simple strategies when optimising for probability of guaranteed success. Considering a larger search space of strategies is only necessary in settings with complex conflict.