Optimal Simple Strategies for Persuasion

Argument dialogues provide a principled way of structuring rational interactions between participants (be they human or machine), each arguing over the validity of certain claims, with each agent aiming for an outcome that achieves their dialogue goal (e.g., to persuade the other participant to accept their point of view [8], or to reach agreement on an action to perform [2]). Achievement of an agent’s dialogue goal typically depends on both the arguments that the agent chooses to make during the dialogue, determined by its strategy, and the arguments asserted by its interlocutor. The strategising agent— the proponent—thus has the difficult problem of having to consider not only which arguments to assert but also the possible responses of its opponent. This problem is compounded since the opponent may exploit knowledge inferred from those arguments asserted by the proponent to construct new arguments. Hence, the proponent must take care not to divulge information that is advantageous to its opponent. The important challenge of how to generate strategies for such a proponent has not been widely explored [10]. Notable exceptions are the work of Hadoux et al. [7], which employs mixed observability Markov decision processes to generate optimal policies for the proponent to follow; the work of Rienstra et al. [9], which applies a variant of the minimax algorithm to determine an effective proponent strategy; and the work of Black et al. [3], which employs heuristic planning techniques to determine an optimal proponent strategy for a simple asymmetric persuasion setting. We highlight two types of uncertainty in the strategic argumentation problem: uncertainty over the arguments initially know to the opponent, captured by the opponent model, and uncertainty over how the opponent chooses to argue, given their initial knowledge base. Both [9] and [3] deal with uncertain models of the opponents initial knowledge base, where the opponent’s strategy is known (i.e., optimal [9] or deterministic [3]); while [7] considers the case in which we have only a stochastic model of how the opponent will behave. The key novelties in our approach are that we deal with both of these types of uncertainty simultaneously the former through use of an uncertain opponent model and the latter by generating conformant strategies, that is strategies that are effective regardless of the opponent strategy. Further, our work is the first to generate strategies in a setting where the opponent may exploit information obtained during the dialogue to construct arguments unknown to it at the start of the dialogue, necessitating more cautious strategies. 2 Strategic Argumentation Problem