Deep Forward-Backward SDEs for Min-max Control Machine Learning

This paper presents a novel approach to numerically solve stochastic differential games for nonlinear systems. The proposed approach relies on the nonlinear Feynman-Kac theorem that establishes a connection between parabolic deterministic partial differential equations and forward-backward stochastic differential equations. Using this theorem the Hamilton-Jacobi-Isaacs partial differential equation associated with differential games is represented by a system of forward-backward stochastic differential equations. Numerical solution of the aforementioned system of stochastic differential equations is performed using importance sampling and a Long-Short Term Memory recurrent neural network, which is trained in an offline fashion. The resulting algorithm is tested on two example systems in simulation and compared against the standard risk neutral stochastic optimal control formulations.

Reinforcement Learning for Continuous Stochastic Control Problems

Neural Information Processing Systems

Here we sudy the continuous time, continuous state-spacestochastic case, which covers a wide variety of control problems including target, viability, optimization problems (see [FS93], [KP95])}or which a formalism is the following.

Reinforcement learning - Scholarpedia


Reinforcement learning (RL) is learning by interacting with an environment. An RL agent learns from the consequences of its actions, rather than from being explicitly taught and it selects its actions on basis of its past experiences (exploitation) and also by new choices (exploration), which is essentially trial and error learning. The reinforcement signal that the RL-agent receives is a numerical reward, which encodes the success of an action's outcome, and the agent seeks to learn to select actions that maximize the accumulated reward over time. In general we are following Marr's approach (Marr et al 1982, later re-introduced by Gurney et al 2004) by introducing different levels: the algorithmic, the mechanistic and the implementation level. The best studied case is when RL can be formulated as class of Markov Decision Problems (MDP). The agent can visit a finite number of states and in visiting a state, a numerical reward will be collected, where negative numbers may represent punishments.

Stochastic modified equations and adaptive stochastic gradient algorithms Machine Learning

We develop the method of stochastic modified equations (SME), in which stochastic gradient algorithms are approximated in the weak sense by continuous-time stochastic differential equations. We exploit the continuous formulation together with optimal control theory to derive novel adaptive hyper-parameter adjustment policies. Our algorithms have competitive performance with the added benefit of being robust to varying models and datasets. This provides a general methodology for the analysis and design of stochastic gradient algorithms.

Improving Policies without Measuring Merits

Neural Information Processing Systems

Performing policy iteration in dynamic programming should only require knowledge of relative rather than absolute measures of the utility of actions (Werbos, 1991) - what Baird (1993) calls the advantages ofactions at states. Nevertheless, most existing methods in dynamic programming (including Baird's) compute some form of absolute utility function. For smooth problems, advantages satisfy two differential consistency conditions (including the requirement that they be free of curl), and we show that enforcing these can lead to appropriate policy improvement solely in terms of advantages. 1 Introd uction In deciding how to change a policy at a state, an agent only needs to know the differences (called advantages) between the total return based on taking each action a for one step and then following the policy forever after, and the total return based on always following the policy (the conventional value of the state under the policy). The advantages are like differentials - they do not depend on the local levels of the total return. Indeed, Werbos (1991) defined Dual Heuristic Programming (DHP), using these facts, learning the derivatives of these total returns with respect to the state.