Stochastic differential games (Q2543740): Difference between revisions

Consider a stochastic differential system of \(m\) equations \[ d\xi(t)=f(t,\xi(t), y_1,\ldots, y_N)\,dt+ \sigma(t, \xi(t))\,dw(t),\quad \xi(s) =x_0, \] where the player \(y_i\) chooses a control function with values in a control set \(Y_i\). Denote by \(\tau\) the exit time of \(\xi(t)\) from a cylinder \(\{s <t\leq T\}\times\Omega\) \((\Omega\) a bounded domain in \(\mathbb R^n\)). The goal of \(y_i\) is to minimize the cost \[ J_i(y_i\ldots, y_N)=E\{\int_s^\tau h_i(t,\xi(t), y_1,\ldots, y_N)\,dt + g_i (\tau, \xi(\tau))\}. \] In a setting of perfect observation the controls used by \(y_i\) are those of pure strategies, i.e., \(y_i = y_i (t,\xi(t))\). It is then proved that an equilibrium strategy exists. In the case of partial observation a more general concept of a strategy is introduced, involving mappings between spaces of controls. In the special case of zero-sum 2-person game it is proved that the value and a saddle point exist. The proofs employ techniques of parabolic partial differential equations.