ADPRL 2014 - IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning (ADPRL 2014)

Adaptive (or Approximate) dynamic programming (ADP) is a general and effective approach for solving optimal control problems by adapting to uncertain environments over time. ADP optimizes a user-defined cost function with respect to an adaptive control law, conditioned on prior knowledge of the system, and its state, in the presence of system uncertainties. A numerical search over the present value of the control minimizes a nonlinear cost function forward-in-time providing a basis for real-time, approximate optimal control. The ability to improve performance over time subject to new or unexplored objectives or dynamics has made ADP an attractive approach in a number of application domains including optimal control and estimation, operation research, and computational intelligence. ADP is viewed as a form of reinforcement learning based on an actor-critic architecture that optimizes a user-prescribed value online and obtains the resulting optimal control policy.
Reinforcement learning (RL) algorithms learn to optimize an agent by letting it interact with an environment and learn from its received feedback. The goal of the agent is to optimize its accumulated reward over time, and for this it estimates value functions that predict its future reward intake when executing a particular policy. Reinforcement learning techniques can be combined with many different function approximators and do not assume any a priori knowledge about the environment. An important aspect in RL is that an agent has to explore parts of the environment it does not know well, while at the same time it has to exploit its knowledge to maximize its reward intake. RL techniques have already been applied successfully for many problems such as controlling robots, game playing, elevator control, network routing, and traffic light optimization.
Topics
The symposium topics include, but are not limited to:
Convergence and performance bounds of ADP
Complexity issues in RL and ADP
Statistical learning and RL, PAC bounds for RL
Monte-Carlo and quasi Monte-Carlo methods
Direct policy search, actor-critic methods
Parsimoneous function representation
Adaptive feature discovery
Learning rules and architectures for RL
Sensitivity analysis for policy gradient estimation
Neuroscience and biologically inspired control
Partially observable Markov decision processes
Distributed intelligent systems
Multi-agent RL systems
Multi-level multi-objective optimization for ADPRL
Kernel methods and value function representation
Applications of ADP and RL