Online Learning of Feasible Strategies in Unknown Environments

Define an environment as a set of convex constraint functions that vary arbitrarily over time and consider a cost function that is also convex and arbitrarily varying. Agents that operate in this environment intend to select actions that are feasible for all times while minimizing the cost's time average. Such action is said optimal and can be computed offline if the cost and the environment are known a priori. An online policy is one that depends causally on the cost and the environment. To compare online policies to the optimal offline action define the fit of a trajectory as a vector that integrates the constraint violations over time and its regret as the cost difference with the optimal action accumulated over time. Fit measures the extent to which an online policy succeeds in learning feasible actions while regret measures its success in learning optimal actions. This paper proposes the use of online policies computed from a saddle point controller which are shown to have fit and regret that are either bounded or grow at a sublinear rate. These properties provide an indication that the controller finds trajectories that are feasible and optimal in a relaxed sense. Concepts are illustrated throughout with the problem of a shepherd that wants to stay close to all sheep in a herd. Numerical experiments show that the saddle point controller allows the shepherd to do so.