next up previous
Next: Issues Up: The Reinforcement Learning Previous: The Reinforcement Learning

Equations for Reinforcement Learning

  Utility (over a finite agent lifetime) is defined as the expected sum of the immediate reward and the long-time reward under the best possible policy:


where is the state at time step t, is the immediate reward of executing action in state , N is the number of steps in the lifetime of the agent, and is the reward at time step t.). The operator stands for taking an expectation over all sources of randomness in the system.

Utility (over an infinite agent lifetime) is defined similarly:


To avoid the mathematical awkwardness of infinite sums, we introduce a discount factor, , which counts future rewards less than immediate rewards. This is similar to the compound interest that banks use.

The utility of a state can be defined in terms of the utility of the next state:


This gives a system of equations, called the Bellman Optimality Equations (e.g., Bertsekas [2]), one for each possible state-action pair, the solution to which is the utility function.

The dynamic programming method for solving the Bellman equations is to iterate the following, :


The RL rule for updating the estimated utility is:


where is a small number less than one that determines the rate of change of the estimate. Notice that the second part of Equation 5 is a lot like Equation 3, except that there are no expectation signs E anywhere. See Barto, Bradtke, and Singh [1] for additional information on RL methods.