This talk discusses models featuring the Conjectural Variations concept in dynamic games. In these models, players in repeated interaction form conjectures about the behavior of their opponent, and confront this conjecture to observations. Two classes of results will be presented.
A first approach is to define dynamic games with consistent conjectures. Several possibilities will be presented. The particular case of "feedback consistency" proposed by J.W. Friedman will be studied in more detail.
A second way is to give to the players the possibility to revise their conjectures over time. This leads to learning models. We will also describe several possibilities in this framework, and show how the evolution of conjectures may lead to Pareto-efficient outcomes in the long run.