next up previous
Next: Q-learning Up: Learning an Optimal Policy: Previous: Learning an Optimal Policy:

Adaptive Heuristic Critic and tex2html_wrap_inline2034

 

The adaptive heuristic critic algorithm is an adaptive version of policy iteration [9] in which the value-function computation is no longer implemented by solving a set of linear equations, but is instead computed by an algorithm called TD(0). A block diagram for this approach is given in Figure 4. It consists of two components: a critic (labeled AHC), and a reinforcement-learning component (labeled RL). The reinforcement-learning component can be an instance of any of the k-armed bandit algorithms, modified to deal with multiple states and non-stationary rewards. But instead of acting to maximize instantaneous reward, it will be acting to maximize the heuristic value, v, that is computed by the critic. The critic uses the real external reinforcement signal to learn to map states to their expected discounted values given that the policy being executed is the one currently instantiated in the RL component.

   figure277
Figure 4: Architecture for the adaptive heuristic critic.

We can see the analogy with modified policy iteration if we imagine these components working in alternation. The policy tex2html_wrap_inline1748 implemented by RL is fixed and the critic learns the value function tex2html_wrap_inline2010 for that policy. Now we fix the critic and let the RL component learn a new policy tex2html_wrap_inline1996 that maximizes the new value function, and so on. In most implementations, however, both components operate simultaneously. Only the alternating implementation can be guaranteed to converge to the optimal policy, under appropriate conditions. Williams and Baird explored the convergence properties of a class of AHC-related algorithms they call ``incremental variants of policy iteration'' [133].

It remains to explain how the critic can learn the value of a policy. We define tex2html_wrap_inline2048 to be an experience tuple summarizing a single transition in the environment. Here s is the agent's state before the transition, a is its choice of action, r the instantaneous reward it receives, and s' its resulting state. The value of a policy is learned using Sutton's TD(0) algorithm [115] which uses the update rule

displaymath2026

Whenever a state s is visited, its estimated value is updated to be closer to tex2html_wrap_inline2062 , since r is the instantaneous reward received and V(s') is the estimated value of the actually occurring next state. This is analogous to the sample-backup rule from value iteration--the only difference is that the sample is drawn from the real world rather than by simulating a known model. The key idea is that tex2html_wrap_inline2062 is a sample of the value of V(s), and it is more likely to be correct because it incorporates the real r. If the learning rate tex2html_wrap_inline1902 is adjusted properly (it must be slowly decreased) and the policy is held fixed, TD(0) is guaranteed to converge to the optimal value function.

The TD(0) rule as presented above is really an instance of a more general class of algorithms called tex2html_wrap_inline1724 , with tex2html_wrap_inline2082 . TD(0) looks only one step ahead when adjusting value estimates; although it will eventually arrive at the correct answer, it can take quite a while to do so. The general tex2html_wrap_inline1724 rule is similar to the TD(0) rule given above,

displaymath2027

but it is applied to every state according to its eligibility e(u), rather than just to the immediately previous state, s. One version of the eligibility trace is defined to be

displaymath2028

The eligibility of a state s is the degree to which it has been visited in the recent past; when a reinforcement is received, it is used to update all the states that have been recently visited, according to their eligibility. When tex2html_wrap_inline2082 this is equivalent to TD(0). When tex2html_wrap_inline2106 , it is roughly equivalent to updating all the states according to the number of times they were visited by the end of a run. Note that we can update the eligibility online as follows:

displaymath2029

It is computationally more expensive to execute the general tex2html_wrap_inline1724 , though it often converges considerably faster for large tex2html_wrap_inline2110  [30, 32]. There has been some recent work on making the updates more efficient [24] and on changing the definition to make tex2html_wrap_inline1724 more consistent with the certainty-equivalent method [108], which is discussed in Section 5.1.


next up previous
Next: Q-learning Up: Learning an Optimal Policy: Previous: Learning an Optimal Policy:

Leslie Pack Kaelbling
Wed May 1 13:19:13 EDT 1996