“Stochastic approximation” is an oddly named field, that I suspect will be of increasing interest to the machine learning community. I would prefer a name like “stochastic optimization” (which appears to be taken by several other areas already.)

The basic idea of stochastic approximation is that we would like to optimize some function . However, we only have *noisy* access to . The type of problem depends on exactly what we can measure.

# Robbins-Monro

In a **Robbins-Monro** problem, we have access to unbiased estimates of the gradient of . That is, for any , we can generate (by performing an experiment, or running a simulation, or whatever) a random vector , where the expectation of is the true gradient of .

The basic result that Robbins and Monro came up with in 1951 is that can be optimized by taking decreasing steps along the estimated gradients. This general stochastic gradient descent algorithm is:

1. For

2. Get a noisy, unbiased estimate of the gradient:

3. Step along the estimated gradient:

The step sizes need to be a sequence satisfying

1)

2) .

(i.e. the step sizes go to zero, but there is always an “infinite amount of step left”). The sequence for some constants and is the typical sequence satisfying this.

It is critical that the above sequence is guaranteed to find a local minima of only *asymptotically*. I think it is underemphasized in the stochastic approximation literature that this is a weak guarantee, both in theory *and in practice*, since any real application can take only a finite number of steps. Tuning the constants and well is critical, and often difficult. It is very common to simply disregard the theory and use a small fixed step size , again suggesting that the above theory should not be taken too seriously. With a fixed step size, one finds in practice that moves close to a minima, and then bounces around forever, with the size of the area in which it is bouncing depending on the variance of . (If you can do anything to reduce the variance of , do it! (See common random numbers, below, someday.))

More recently, Polyak and Juditsky worked on “iterate averaging”. The idea is to let the step sizes go to zero slower than , but consider the solution to be the average of the over some previous number of steps. (The number of steps that we average over increases with ). This can lead to faster convergence.

# Kiefer-Wolfowitz

In a **Kiefer-Wolfowitz** problem, one does not have access to noisy estimates of , but only itself. In their 1952 paper, they suggested, essentially, estimating the gradient by a finite-difference kind of scheme.

There are two major problems with this. First, to estimate one gradient, we will need to get estimates of , where is the number of dimensions of . This could be very expensive. Second, for any given , the above is a *biased* estimate of the true gradient. One can reduce this bias arbitrarily by making very small, but this causes a catastrophic *variance explosion*: For small , the estimated gradient will be dominated by the random fluctuations in , rather than the fluctuation to the the change of in which we are interested. (The noise is of order ) This variance can be reduced, either again by techniques like common random numbers, or by, e.g. estimating many gradients and averaging them. Basically, for K-W to be practical, one should have a problem with few dimensions, a function that is cheap to evaluate and/or low variance in .

1. For

2. Get a noisy, unbiased estimate of the gradient:

3. Step along the estimated gradient:

The Kiefer-Wolfowitz algorithm is guaranteed to asymptotically converge to a minima of with the sequences and both slowly going to zero. In practice, however, because of the “variance explosion”, above, one almost always just picks a fixed constant . This keeps the variance bounded, at the cost of biasing somewhat the final solution. Interestingly, Flaxman et al. showed that this amounts to minimizing a “smoothed” version of , where the smoothing depends on .

To come:

1.** SPSA**

2. **Infinitesimal Perturbation Analysis**

3. The **Likelihood Ratio** method.

4. **Common Random Numbers**, and **Variance Reduction Methods** more generally

5. Lecun et al.’s heuristics for training neural networks as variance reduction methods.

A standard reference on stochastic approximation is “Stochastic Approximation Algorithms and Applications” by Kushner and Yin. Spall’s “Introduction to Stochastic Search and Optimization” is more applied and more accessible.

Recommended:

On the choice of random directions for stochastic approximation algorithms by James Theiler, Jarod Alper

Some papers on optimization by random directions:

Learning curves for stochastic gradient descent in linear feedforward networks by Justin Werfel and Xiaohui Xie

A fast stochastic Error-Descent Algorithm for Supervised Learning and Optimization by Gert Cauwenberghs