Gradientless Descent: High-Dimensional Zeroth-Order Optimization

  • 2019-11-14 18:58:13
  • Daniel Golovin, John Karro, Greg Kochanski, Chansoo Lee, Xingyou Song, Qiuyi, Zhang
  • 10


Zeroth-order optimization is the process of minimizing an objective $f(x)$,given oracle access to evaluations at adaptively chosen inputs $x$. In thispaper, we present two simple yet powerful GradientLess Descent (GLD) algorithmsthat do not rely on an underlying gradient estimate and are numerically stable.We analyze our algorithm from a novel geometric perspective and present a novelanalysis that shows convergence within an $\epsilon$-ball of the optimum in$O(kQ\log(n)\log(R/\epsilon))$ evaluations, for {\it any monotone transform} ofa smooth and strongly convex objective with latent dimension $k < n$, where theinput dimension is $n$, $R$ is the diameter of the input space and $Q$ is thecondition number. Our rates are the first of its kind to be both 1)poly-logarithmically dependent on dimensionality and 2) invariant undermonotone transformations. We further leverage our geometric perspective to showthat our analysis is optimal. Both monotone invariance and its ability toutilize a low latent dimensionality are key to the empirical success of ouralgorithms, as demonstrated on BBOB and MuJoCo benchmarks.


Quick Read (beta)

Gradientless Descent: High-Dimensional Zeroth-Order Optimization

Daniel Golovin    John Karro    Greg Kochanski    Chansoo Lee    Xingyou Song    Qiuyi (Richard) Zhang Author list in alphabetical order.
Google Brain thanks: {dgg,karro,gpk,chansoo,xingyousong,qiuyiz}

Zeroth-order optimization is the process of minimizing an objective f(x), given oracle access to evaluations at adaptively chosen inputs x. In this paper, we present two simple yet powerful GradientLess Descent (GLD) algorithms that do not rely on an underlying gradient estimate and are numerically stable. We analyze our algorithm from a novel geometric perspective and present a novel analysis that shows convergence within an ϵ-ball of the optimum in O(kQlog(n)log(R/ϵ)) evaluations, for any monotone transform of a smooth and strongly convex objective with latent dimension k<n, where the input dimension is n, R is the diameter of the input space and Q is the condition number. Our rates are the first of its kind to be both 1) poly-logarithmically dependent on dimensionality and 2) invariant under monotone transformations. We further leverage our geometric perspective to show that our analysis is optimal. Both monotone invariance and its ability to utilize a low latent dimensionality are key to the empirical success of our algorithms, as demonstrated on BBOB and MuJoCo benchmarks.

1 Introduction

We consider the problem of zeroth-order optimization (also known as gradient-free optimization, or bandit optimization), where our goal is to minimize an objective function f:n with as few evaluations of f(x) as possible. For many practical and interesting objective functions, gradients are difficult to compute and there is still a need for zeroth-order optimization in applications such as reinforcement learning [MGR18, SHC+17, CRS+18], attacking neural networks [CZS+17, PMG+17], hyperparameter tuning of deep networks [SLA12], and network control [LCC+17].

The standard approach to zeroth-order optimization is, ironically, to estimate the gradients from function values and apply a first-order optimization algorithm [FKM05]. [NS11] analyze this class of algorithms as gradient descent on a Gaussian smoothing of the objective and gives an accelerated O(nQlog((LR2+F)/ϵ)) iteration complexity for an L-Lipschitz convex function with condition number Q and R=x0-x* and F=f(x0)-f(x*). They propose a two-point evaluation scheme that constructs gradient estimates from the difference between function values at two points that are close to each other. This scheme was extended by [DJW+15] for stochastic settings, by [GL13] for nonconvex settings, and by [SHA17] for non-smooth and non-Euclidean norm settings. Since then, first-order techniques such as variance reduction [LKC+18], conditional gradients [BG18], and diagonal preconditioning [MGR18] have been successfully adopted in this setting. This class of algorithms are also known as stochastic search, random search, or (natural) evolutionary strategies and have been augmented with a variety of heuristics, such as the popular CMA-ES [AH05].

These algorithms, however, suffer from high variance due to non-robust local minima or highly non-smooth objectives, which are common in the fields of deep learning and reinforcement learning. [MGR18] notes that gradient variance increases as training progresses due to higher variance in the objective functions, since often parameters must be tuned precisely to achieve reasonable models. Therefore, some attention has shifted into direct search algorithms that usually finds a descent direction u and moves to x+δu, where the step size is not scaled by the function difference.

The first approaches for direct search were based on deterministic approaches with a positive spanning set and date back to the 1950s [BRO58]. Only recently have theoretical bounds surfaced, with [GRV+15] giving an iteration complexity that is a large polynomial of n and [DV16] giving an improved O(n2L2/ϵ). Stochastic approaches tend to have better complexities: [SMG13] uses line search to give a O(nQlog(F/ϵ)) iteration complexity for convex functions with condition number Q and most recently, [GBS+19] uses importance sampling to give a O(nQ¯log(F/ϵ)) complexity for convex functions with average condition number Q¯, assuming access to sampling probabilities. [SMG13] notes that direct search algorithms are invariant under monotone transforms of the objective, a property that might explain their robustness in high-variance settings.

In general, zeroth order optimization suffers an at least linear dependence on input dimension n and recent works have tried to address this limitation when n is large but f(x) admits a low-dimensional structure. Some papers assume that f(x) depends only on k coordinates and [WDB+17] applies Lasso to find the important set of coordinates, whereas [BG18] simply change the step size to achieve an O(k(log(n)/ϵ)2) iteration complexity. Other papers assume more generally that f(x)=g(𝐏𝐀x) only depends on a k-dimensional subspace given by the range of 𝐏𝐀 and [DKC13] apply low-rank approximation to find the low-dimensional subspace while [WZH+13] use random embeddings. [HKY17] assume that f(x) is a sparse collection of k-degree monomials on the Boolean hypercube and apply sparse recovery to achieve a O(nk) runtime bound. We will show that under the case that f(x)=g(𝐏𝐀x), our algorithm will inherently pick up any low-dimensional structure in f(x) and achieve a convergence rate that depends on klog(n). This initial convergence rate survives, even if we perturb f(x)=g(𝐏𝐀x)+h(x), so long as h(x) is sufficiently small.

We will not cover the whole variety of black-box optimization methods, such as Bayesian optimization or genetic algorithms. In general, these methods attempt to solve a broader problem (e.g. multiple optima), have weaker theoretical guarantees and may require substantial computation at each step: e.g. Bayesian optimization generally has theoretical iteration complexities that grow exponentially in dimension, and CMA-ES lacks provable complexity bounds beyond convex quadratic functions. In addition to the slow runtime and weaker guarantees, Bayesian optimization assumes the success of an inner optimization loop of the acquisition function. This inner optimization is often implemented with many iterations of a simpler zeroth-order methods, justifying the need to understand gradient-less descent algorithms within its own context.

1.1 Our contributions

In this paper, we present GradientLess Descent (GLD), a class of truly gradient-free algorithms (also known as direct search algorithms) that are parameter free and provably fast. Our algorithms are based on a simple intuition: for well-conditioned functions, if we start from a point and take a small step in a randomly chosen direction, there is a significant probability that we will reduce the objective function value. We present a novel analysis that relies on facts in high dimensional geometry and can thus be viewed as a geometric analysis of gradient-free algorithms, recovering the standard convergence rates and step sizes. Specifically, we show that if the step size is on the order of O(1n), we can guarantee an expected decrease of 1-Ω(1n) in the optimality gap, based on geometric properties of the sublevel sets of a smooth and strongly convex function.

Our results are invariant under monotone transformations of the objective function, thus our convergence results also hold for a large class of non-convex functions that are a subclass of quasi-convex functions. Specifically, note that monotone transformations of convex functions are not necessarily convex. However, a monotone transformation of a convex function is always quasi-convex. The maximization of quasi-concave utility functions, which is equivalent to the minimization of quasi-convex functions, is an important topic of study in economics (e.g. [AE61]).

Intuition suggests that the step-size dependence on dimensionality can be improved when f(x) admits a low-dimensional structure. With a careful choice of sampling distribution we can show that if f(x)=g(𝐏𝐀x), where 𝐏𝐀 is a rank k matrix, then our step size can be on the order of O(1k) as our optimization behavior is preserved under projections. We call this property affine-invariance and show that the number of function evaluations needed for convergence depends logarithmically on n. Unlike most previous algorithms in the high-dimensional setting, no expensive sparse recovery or subspace finding methods are needed. Furthermore, by novel perturbation arguments, we show that our fast convergence rates are robust and holds even under the more realistic assumption when f(x)=g(𝐏𝐀x)+h(x) with h(x) being sufficiently small.

Theorem 1 (Convergence of GLD: Informal Restatement of Theorem 7 and Theorem 14).

Let f(x) be any monotone transform of a convex function with condition number Q and R=x0-x*. Let y be a sample from an appropriate distribution centered at x. Then, with constant probability,


Therefore, we can find xT such that xT-x*ϵ after T=O~(nQlog(R/ϵ)) function evaluations. Furthermore, for functions f(x)=g(PAx)+h(x) with rank k matrix PA and sufficiently small h(x), we only require O~(kQlog(n)log(R/ϵ)) evaluations.

Another advantage of our non-standard geometric analysis is that it allows us to deduce that our rates are optimal with a matching lower bound (up to logarithmic factors), presenting theoretical evidence that gradient-free inherently requires Ω(nQ) function evaluations to converge. While gradient-estimation algorithms can achieve a better theoretical iteration complexity of O(nQ), they lack the monotone and affine invariance properties. Empirically, we see that invariance properties are important to successful optimization, as validated by experiments on synthetic BBOB and MuJoCo benchmarks that show the competitiveness of GLD against standard optimization procedures.

Table 1: Comparison of zeroth order optimization for well-conditioned convex functions where R=x0-x* and F=f(x0)-f(x*). ‘Monotone’ column indicates the invariance under monotone transformations (Definition 4). ‘k-Sparse’ and ‘k-Affine’ columns indicate that iteration complexity is poly(k, log(n)) when f(x) depends only on a k-sparse subset of coordinates or on a rank-k affine subspace.
Algorithm Iteration Complexity Monotone k-Sparse k-Affine
[NS11] nlog((R2+F)/ϵ) No No No
[BG18] nlog(F/ϵ) No Yes No
[SMG13] nlog(F/ϵ) Yes No No
[GBS+19] nlog(F/ϵ) Yes No No
This paper (GLD) nlog(R/ϵ) Yes Yes Yes

2 Preliminaries

We first define a few notations for the rest of the paper. Let 𝒳 be a compact subset of n and let denote the Euclidean norm. The diameter of 𝒳, denoted 𝒳=maxx,x𝒳x-x, is the maximum distance between elements in 𝒳. Let f:𝒳 be a real-valued function which attains its minimum at x*. We use f(𝒳)={f(x):x𝒳} to denote the image of f on a subset 𝒳 of n, and (c,r)={xn:c-xr} to denote the ball of radius r centered at c.

Definition 2.

The level set of f at point xX is Lc(f)={yX:f(y)=f(x)}. The sub-level set of f at point xX is Lc(f)={yX:f(y)f(x)}. When the function f is clear from the context, we omit it.

Definition 3.

We say that f is α-strongly convex for α>0 if f(y)f(x)+f(x),y-x+α2y-x2 for all x,yX and β-smooth for β>0 if f(y)f(x)+f(x),y-x+β2y-x2 for all x,yX.

Definition 4.

We say that gf is a monotone transformation of f if g:f(X)R is a monotonically (and strictly) increasing function.

Monotone transformations preserve the level sets of a function in the sense that x(f)=x(gf). Because our algorithms depend only on the level set properties, our results generalize to any monotone transformation of a strongly convex and strongly smooth function. This leads to our extended notion of condition number.

Definition 5.

A function f has condition number Q1 if it is the minimum ratio β/α over all functions g such that f is a monotone transformation of g and g is α-strongly convex and β smooth.

When we work with low rank extensions of f, we only care about the condition number of f within a rank k subspace. Indeed, if f only varies along a rank k subspace, then it has a strong convexity value of 0, making its condition number undefined. If f is α-strongly convex and β-smooth, then its Hessian matrix always has eigenvalues bounded between α and β. Therefore, we need a notion of a projected condition number. Let 𝐀d×k be some orthonormal matrix and let 𝐏𝐀=𝐀𝐀 be the projection matrix onto the column space of 𝐀.

Definition 6.

For some orthonormal ARd×k with d>k, a function f has condition number restricted to A, Q(A)1, if it is the minimum ratio β/α over all functions g such that f is a monotone transformation of g and h(y)=g(Ay) is α-strongly convex and β smooth.

3 Analysis of Descent Steps

The GLD template can be summarized as follows: given a sampling distribution 𝒟, we start at x0 and in iteration t, we choose a scalar radii rt and we sample yt from a distribution rt𝒟 centered around xt, where rt provides the scaling of 𝒟. Then, if f(xt+1)<xt, we update xt+1=yt; otherwise, we set xt+1=xt. The analysis of GLD follows from the main observation that the sub-level set of a monotone transformation of a strongly convex and strongly smooth function contains a ball of sufficiently large radius tangent to the level set (Lemma 15). In this section, we show that this property, combined with facts of high-dimensional geometry, implies that moving in a random direction from any point has a good chance of significantly improving the objective.

As we mentioned before, the key to fast convergence is the careful choice of step sizes, which we describe in Theorem 7. The intuition here is that we would like to take as large steps as possible while keeping the probability of improving the objective function reasonably high, so by insights in high-dimensional geometry, we choose a step size of Θ(1/n). Also, we show that if f(x) admits a latent rank-k structure, then this step size can be increased to Θ(1/k) and is therefore only dependent on the latent dimensionality of f(x), allowing for fast high-dimensional optimization. Lastly, our geometric understanding allows us to show that our convergence rates are optimal with a matching lower bound. Without loss of generality, this section assumes that f(x) is strongly convex and smooth with condition number Q.

3.1 Step Size

Theorem 7.

For any x such that 35Qx-x*[C1,C2], we can find integers 0k1,k2<logC2C1 such that if r=2k1C1 or r=2-k2C2, then a random sample y from uniform distribution over Bx=B(x,rn) satisfies


with probability at least 14.

Proving the above theorem requires the following lemma about the intersection of balls in high dimensions and it is proved in the appendix.

Lemma 8.

Let B1 and B2 be two balls in Rn of radii r1 and r2 respectively. Let be the distance between the centers. If r1[2n,n] and r2-4n, then


where cn is a dimension-dependent constant that is lower bounded by 14 at n=1.

3.2 Gaussian Sampling and Low Rank Structure

A direct application of Lemma 8 seems to imply that uniform sampling of a high-dimensional ball is necessary. Upon further inspection, this can be easily replaced with a much simpler Gaussian sampling procedure that concentrates the mass close to the surface to the ball. This procedure lends itself to better analysis when f(x) admits a latent low-dimensional structure since any affine projection of a Gaussian is still Gaussian.

Lemma 9.

Let B1 and B2 be two balls in Rn of radii r1 and r2 respectively. Let be the distance between the centers. If r1[2n,n] and r2-n and X=(X1,,Xn) are independent Gaussians with mean centered at the center of B1 and variance r12n, then


where c is a dimension-independent constant.

Assume that there exists some rank k projection matrix 𝐏𝐀 such that f(x)=g(𝐏𝐀x), where k is much smaller than n. Because Gaussians projected on a k-dimensional subspace are still Gaussians, we show that our algorithm has a dimension dependence on k. We let Qg(𝐀) be the condition number of g restricted to the subspace 𝐀 that drives the dominant changes in f(x).

Theorem 10.

Let f(x)=g(PAx) for some unknown rank k matrix PA with k<n and suppose 35QPA(x-x*)[C1,C2] for some numbers C1,C2R+. Then, there exist integers 0k1,k2<logC2C1 such that if r=2k1C1 or r=2-k2C2, then a random sample y from a Gaussian distribution N(x,r2kI) satisfies


with constant probability.

Note that the speed-up in progress is due to the fact that we can now tolerate the larger sampling radius of Ω(1/k), while maintaining a high probability of making progress. If k is unknown, we can simply use binary search to find the correct radius with an extra factor of log(n) in our runtime.

The low-rank assumption is too restrictive to be realistic; however, our fast rates still hold, at least for the early stages of the optimization, even if we assume that f(x)=g(𝐏𝐀x)+h(x) and |h(x)|δ is a full-rank function that is bounded by δ. In this setting, we can show that convergence remains fast, at least until the optimality gap approaches δ.

Theorem 11.

Let f(x)=g(PAx)+h(x) for some unknown rank k matrix PA with k<n where g,h are convex and |h|δ. Suppose 35QPAx-z*[C1,C2] for some numbers C1,C2R+ where z* minimizes g(z). Then, there exist integers 0k1,k2<logC2C1 such that if r=2k1C1 or r=2-k2C2, then a random sample y from a Gaussian distribution N(x,r2kI) satisfies


with constant probability whenever f(x)-f(x*)60δkQg(A).

3.3 Lower Bounds

We show that our upper bounds given in the previous section are tight up to logarithmic factors for any symmetric sampling distribution 𝒟. These lower bounds are easily derived from our geometric perspective as we show that a sampling distribution with a large radius gives an extremely low probability of intersection with the desired sub-level set. Therefore, while gradient-approximation algorithms can be accelerated to achieve a runtime that depends on the square-root of the condition number Q, gradient-less methods that rely on random sampling are likely unable to be accelerated according to our lower bound. However, we emphasize that monotone invariance allows these results to apply to a broader class of objective functions, beyond smooth and convex, so the results can be useful in practice despite the seemingly worse theoretical bounds.

Theorem 12.

Let y=x+v, where v is a random sample from rD for some radius r>0 and D is standard Gaussian or any rotationally symmetric distribution. Then, there exist a region X with positive measure such that for any xX,


with probability at least 1-1𝑝𝑜𝑙𝑦(nQ).

4 Gradientless Algorithms

In this section, we present two algorithms that follow the same Gradientless Descent (GLD) template: GLD-Search and GLD-Fast, with the latter being an optimized version of the former when an upper bound on the condition number of a function is known. For both algorithms, since they are monotone-invariant, we appeal to the previous section to derive fast convergence rates for any monotone transform of convex f(x) with good condition number. We show the efficacy of both algorithms experimentally in the Experiments section.

4.1 Gradientless Descent with Binary Search

Although the sampling distribution 𝒟 is fixed, we have a choice of radii for each iteration of the algorithm. We can apply a binary search procedure to ensure progress. The most straightforward version of our algorithm is thus with a naive binary sweep across an interval in [r,R] that is unchanged throughout the algorithm. This allows us to give convergence guarantees without previous knowledge of the condition number at a cost of an extra factor of log(n/ϵ).

Input : function: f:n, T+: number of iterations, x0: starting point,
𝒟: sampling distribution, R: maximum search radius, r: minimum search radius
1 Set K=log(R/r)
2 for t=0,,T do
3        Ball Sampling Trial i:
4        for k = 0, …, K do
5              Set ri,k=2-kR.
6              Sample vi,kri,k𝒟.
8        end for
9       Update: xt+1=argmink{f(y)|y=xt,y=xt+vi,k}
11 end for
return xt
Algorithm 1 Gradientless Descent with Binary Search (GLD-Search)
Theorem 13.

Let x0 be any starting point and f a blackbox function with condition number Q. Running Algorithm 1 with r=ϵn,R=X and D=N(0,I) as a standard Gaussian returns a point xT such that xT-x*2Q3/2ϵ after O(nQlog(nX/ϵ)2) function evaluations with high probability.

Furthermore, if f(x)=g(PAx) admits a low-rank structure with PA a rank k matrix, then we only require O(kQg(A)log(nX/ϵ)2) function evaluations to guarantee PA(xT-x*)ϵ. This holds analogously even if f(x)=g(PAx)+h(x) is almost low-rank where |h|δ and ϵ>60δkQg(A).

4.2 Gradientless Descent with Fast Binary Search

GLD-Search (Algorithm 1) uses a naive lower and upper bound for the search radius xt-x*, which incurs an extra factor of log(1/ϵ) in the runtime bound. In GLD-Fast, we remove this extra factor dependence on log(1/ϵ) by drastically reducing the range of the binary search. This is done by exploiting the assumption that f has a good condition number upper bound Q^ and by slowly halfing the diameter of the search space every few iterations since we expect xtx* as t.

Input : function f:n, T+: number of iterations, x0: starting point,
𝒟: sampling distribution, R: diameter of search space, Q: condition number bound
1 Set K=log(4Q), H=nQlog(Q)
2 for t=1,,T do
3        Set R=R/2 when T0modH (every H iterations).
4        Ball Sampling Trial i:
5        for k = -K, …, 0, …, K do
6              Set ri,k=2-kR.
7              Sample vi,kri,k𝒟.
9        end for
10       Update: xt+1=argmini{f(y)|y=xt,y=xt+vi}
12 end for
return xt
Algorithm 2 Gradientless Descent with Fast Binary Search (GLD-Fast)
Theorem 14.

Let x0 be any starting point and f a blackbox function with condition number upper bounded by Q. Running Algorithm 2 with suitable parameters returns a point xT such that f(xT)-f(x*)ϵ after O(nQlog2(Q)log(X/ϵ)) function evaluations with high probability.

Furthermore, if f(x)=g(PAx) admits a low-rank structure with PA a rank k matrix, then we only require O(kQg(A)log(n)log2(Qg(A))log(X/ϵ)) function evaluations to guarantee PA(xT-x*)ϵ. This holds analogously even if f(x)=g(PAx)+h(x) is almost low-rank where |h|δ and ϵ>60δkQg(A).

5 Experiments

We tested GLD algorithms on a simple class of objective functions and compare it to Accelerated Random Search (ARS) by [NS11], which has linear convergence guarantees on strongly convex and strongly smooth functions. To our knowledge, ARS makes the weakest assumption among the zeroth-order algorithms that have linear convergence guarantees and perform only a constant order of operations per iteration. Our main conclusion is that GLD-Fast is comparable to ARS and tends to achieve a reasonably low error much faster than ARS in high dimensions (50). In low dimensions, GLD-Search is competitive with GLD-Fast and ARS though it requires no information about the function.

We let Hα,β,nn×n be a diagonal matrix with its i-th diagonal equal to α+(β-α)i-1n-1. In simple words, its diagonal elements form an evenly space sequence of numbers from α to β. Our objective function is then fα,β,n:n as fα,β,n(x)=12xHα,β,nx, which is α-strongly convex and β-strongly smooth. We always use the same starting point x=1n(1,,1), which requires X=Q for our algorithms. We plot the optimality gap f(bt)-f(x*) against the number of function evaluations, where bt is the best point observed so far after t evaluations. Although all tested algorithms are stochastic, they have a low variance on the objective functions that we use; hence we average the results over 10 runs and omit the error bars in the plots.

We ran experiments on f1,8,n with imperfect curvature information α^ and β^ (see Figure 3 in appendix). GLD-Search is independent of the condition number. GLD-Fast takes only one parameter, which is the upper bound on the condition number; if approximation factor is z, then we pass 8z as the upper bound. ARS requires both strong convexity and smoothness parameters. We test three different distributions of the approximation error; when the approximation factor is z, then ARS-alpha gets (α/z,β), ARS-beta gets (α,zβ), and ARS-even gets (α/z,zβ) as input. GLD-Fast is more robust and faster than ARS when the condition number is over-approximated. When the condition number is underestimated, GLD-Fast still steadily converges.

5.1 Monotone Transformations

Figure 1: The average optimality gap on a quadratic objective function that is strongly convex and smooth objective (top); and its monotone transformation (bottom). Further experiments on non-convex BBOB functions show similar behavior and are in the appendix.

In Figure 1, we ran experiments on f1,8,n for different settings of dimensionality n, and its monotone transformation with g(y)=-exp(-y). For this experiment, we assume a perfect oracle for the strong convexity and smoothness parameters of f. The convergence of GLD is totally unaffected by the monotone transformation. For the low-dimension cases of a transformed function (bottom half of the figure), we note that there are inflection points in the convergence curve of ARS. This means that ARS initially struggles to gain momentum and then struggles to stop the momentum when it gets close to the optimum. Another observation is that unlike ARS that needs to build up momentum, GLD-Fast starts from a large radius and therefore achieves a reasonably low error much faster than ARS, especially in higher dimensions.

5.2 BBOB Benchmarks

To show that practicality of GLD on practical and non-convex settings, we also test GLD algorithms on a variety of BlackBox Optimization Benchmarking (BBOB) functions [HFR+09]. For each function, the optima is known and we use the log optimality gap as a measure of competance. Because each function can exhibit varying forms of non-smoothness and convexity, all algorithms are ran with a smoothness constant of 10 and a strong convexity constant of 0.1. All other setup details are same as before, such as using a fixed starting point.

The plots, given in Appendix C, underscore the superior performance of GLD algorithms on various BBOB functions, demonstrating that GLD can successfully optimize a diverse set of functions even without explicit knowledge of condition number. We note that BBOB functions are far from convex and smooth, many exhibiting high conditioning, multi-modal valleys, and weak global structure. Due to our radius search produce, our algorithm appears more robust to non-ideal settings with non-convexity and ill conditioning. As expected, we note that GLD-Fast tend to outperform GLD-Search, especially as the dimension increases, matching our theoretical understanding of GLD.

5.3 Mujoco Control Benchmarks and Affine Transformations

We also ran experiments on the Mujoco benchmarks with varying architectures, both linear and nonlinear. This demonstrates the viability of our approach even in the non-convex, high dimensional setting. We note that however, unlike e.g. ES which uses all queries to form a gradient direction, our algorithm removes queries which produce less reward than using the current arg-max, which can be an information handicap. Nevertheless, we see that our algorithm still achieves competitive performance on the maximum reward. We used a horizon of 1000 for all experiments.

Table 2: Final rewards by GLD with linear (L) and deep (H41) policies on Mujoco Benchmarks show that GLD is competitive. We apply an affine projection on HalfCheetah to test affine invariance. We use the reward threshold found from [MGR18] with Reacher’s threshold [SWD+17] for a reasonable baseline.
Env. Arch Rew. at (104, 105, Max) Queries Rew. Thresh.
HalfCheetah-v1 L 3799, 3903, 4064 3430
HalfCheetah-v1, Proj-200 L 1594 , 3509 , 4342 3430
HalfCheetah-v1 H41 2741, 3074, 3392 3430
Hopper-v1 L 1017, 3359, 3375 3120
Hopper-v1 H41 2708, 3370, 3566 3120
Reacher-v1 L -70, -5, -4 -10
Reacher-v1 H41 -231, -17, -15 -10
Swimmer-v1 L 365, 369 , 369 325
Swimmer-v1 H41 353, 369, 369 325
Walker2d-v1 L 1027 , 2201, 2201 4390
Walker2d-v1 H41 1630, 1963, 2146 4390

We further tested the affine invariance of GLD on the policy parameters from using Gaussian ball sampling, under the HalfCheetah benchmark by projecting the state s of the MDP with linear policy to a higher dimensional state Ws, using a matrix multiplication with an orthonormal W. Specifically, in this setting, for a linear policy parametrized by matrix K, the objective function is thus J(KW) where πK(Ws)=KWs. Note that when projecting into a high dimension, there is a slowdown factor of logdnewdold where dnew,dold are the new high dimension and previous base dimension, respectively, due to the binary search in our algorithm on a higher dimensional space. For our HalfCheetah case, we projected the 17 base dimension to a 200-length dimension, which suggests that the slowdown factor is a factor log200173.5. This can be shown in our plots in the appendix (Figure 15).

6 Conclusion

We introduced GLD, a robust zeroth-order optimization algorithm that is simple, efficient, and we show strong theoretical convergence bounds via our novel geometric analysis. As demonstrated by our experiments on BBOB and MuJoCo benchmarks, GLD performs very robustly even in the non-convex setting and its monotone and affine invariance properties give theoretical insight on its practical efficiency.

GLD is very flexible and allows easy modifications. For example, it could use momentum terms to keep moving in the same direction that improved the objective, or sample from adaptively chosen ellipsoids similarly to adaptive gradient methods. [DHS11, MS10]. Just as one may decay or adaptively vary learning rates for gradient descent, one might use a similar change the distribution from which the ball-sampling radii are chosen, perhaps shrinking the minimum radius as the algorithm progresses, or concentrating more probability mass on smaller radii.

Likewise, GLD could be combined with random restarts or other restart policies developed for gradient descent. Analogously to adaptive per–coordinate learning rates [DHS11, MS10], one could adaptively change the shape of the balls being sampled into ellipsoids with various length-scale factors. Arbitrary combinations of the above variants are also possible.


  • [AE61] K. J. Arrow and A. C. Enthoven (1961) Quasi-concave programming. Econometrica: Journal of the Econometric Society, pp. 779–800. Cited by: §1.1.
  • [AH05] A. Auger and N. Hansen (2005) A restart cma evolution strategy with increasing population size. In Evolutionary Computation, 2005. The 2005 IEEE Congress on, Vol. 2, pp. 1769–1776. Cited by: §1.
  • [BG18] K. Balasubramanian and S. Ghadimi (2018) Zeroth-order (non)-convex stochastic optimization via conditional gradient and gradient updates. In Advances in Neural Information Processing Systems, pp. 3455–3464. Cited by: Table 1, §1.
  • [BRO58] S. H. Brooks (1958) A discussion of random methods for seeking maxima. Operations research 6 (2), pp. 244–251. Cited by: §1.
  • [CZS+17] P. Chen, H. Zhang, Y. Sharma, J. Yi, and C. Hsieh (2017) Zoo: zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pp. 15–26. Cited by: §1.
  • [CRS+18] K. Choromanski, M. Rowland, V. Sindhwani, R. E. Turner, and A. Weller (2018) Structured evolution with compact architectures for scalable policy optimization. arXiv preprint arXiv:1804.02395. Cited by: §1.
  • [DKC13] J. Djolonga, A. Krause, and V. Cevher (2013) High-dimensional gaussian process bandits. In Advances in Neural Information Processing Systems, pp. 1025–1033. Cited by: §1.
  • [DV16] M. Dodangeh and L. N. Vicente (2016) Worst case complexity of direct search under convexity. Mathematical Programming 155 (1-2), pp. 307–332. Cited by: §1.
  • [DJW+15] J. C. Duchi, M. I. Jordan, M. J. Wainwright, and A. Wibisono (2015) Optimal rates for zero-order convex optimization: the power of two function evaluations. IEEE Transactions on Information Theory 61 (5), pp. 2788–2806. Cited by: §1.
  • [DHS11] J. Duchi, E. Hazan, and Y. Singer (2011) Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research 12 (Jul), pp. 2121–2159. Cited by: §6.
  • [FKM05] A. D. Flaxman, A. T. Kalai, and H. B. McMahan (2005) Online convex optimization in the bandit setting: gradient descent without a gradient. In Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms, pp. 385–394. Cited by: §1.
  • [GL13] S. Ghadimi and G. Lan (2013) Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization 23 (4), pp. 2341–2368. Cited by: §1.
  • [GBS+19] E. Gorbunov, A. Bibi, O. Sener, E. H. Bergou, and P. Richtárik (2019) A stochastic derivative free optimization method with momentum. arXiv preprint arXiv:1905.13278. Cited by: Table 1, §1.
  • [GRV+15] S. Gratton, C. W. Royer, L. N. Vicente, and Z. Zhang (2015) Direct search based on probabilistic descent. SIAM Journal on Optimization 25 (3), pp. 1515–1541. Cited by: §1.
  • [HFR+09] N. Hansen, S. Finck, R. Ros, and A. Auger (2009) Real-Parameter Black-Box Optimization Benchmarking 2009: Noiseless Functions Definitions. Research Report Technical Report RR-6829, INRIA. Cited by: §5.2.
  • [HKY17] E. Hazan, A. Klivans, and Y. Yuan (2017) Hyperparameter optimization: a spectral approach. arXiv preprint arXiv:1706.00764. Cited by: §1.
  • [LI11] S. Li (2011) Concise formulas for the area and volume of a hyperspherical cap. Asian Journal of Mathematics and Statistics 4 (1), pp. 66–70. Cited by: Appendix A.
  • [LCC+17] S. Liu, J. Chen, P. Chen, and A. O. Hero (2017) Zeroth-order online alternating direction method of multipliers: convergence analysis and applications. arXiv preprint arXiv:1710.07804. Cited by: §1.
  • [LKC+18] S. Liu, B. Kailkhura, P. Chen, P. Ting, S. Chang, and L. Amini (2018) Zeroth-order stochastic variance reduction for nonconvex optimization. In Advances in Neural Information Processing Systems, pp. 3727–3737. Cited by: §1.
  • [MGR18] H. Mania, A. Guy, and B. Recht (2018) Simple random search provides a competitive approach to reinforcement learning. arXiv preprint arXiv:1803.07055. Cited by: §1, Table 2.
  • [MS10] H. B. McMahan and M. J. Streeter (2010) Adaptive bound optimization for online convex optimization. In COLT 2010 - The 23rd Conference on Learning Theory, Haifa, Israel, June 27-29, 2010, pp. 244–256. Cited by: §6.
  • [NS11] Y. Nesterov and V. Spokoiny (2011) Random gradient-free minimization of convex functions. Technical report Université catholique de Louvain, Center for Operations Research and Econometrics (CORE). Cited by: Table 1, §1, §5.
  • [PMG+17] N. Papernot, P. McDaniel, I. Goodfellow, S. Jha, Z. B. Celik, and A. Swami (2017) Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pp. 506–519. Cited by: §1.
  • [SHC+17] T. Salimans, J. Ho, X. Chen, S. Sidor, and I. Sutskever (2017) Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint arXiv:1703.03864. Cited by: §1.
  • [SWD+17] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. CoRR abs/1707.06347. External Links: Link, 1707.06347 Cited by: Table 2.
  • [SHA17] O. Shamir (2017) An optimal algorithm for bandit and zero-order convex optimization with two-point feedback. Journal of Machine Learning Research 18 (52), pp. 1–11. Cited by: §1.
  • [SLA12] J. Snoek, H. Larochelle, and R. P. Adams (2012) Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, pp. 2951–2959. Cited by: §1.
  • [SMG13] S. U. Stich, C. L. Muller, and B. Gartner (2013) Optimization of convex functions with random pursuit. SIAM Journal on Optimization 23 (2), pp. 1284–1309. Cited by: Table 1, §1.
  • [WDB+17] Y. Wang, S. Du, S. Balakrishnan, and A. Singh (2017) Stochastic zeroth-order optimization in high dimensions. arXiv preprint arXiv:1710.10551. Cited by: §1.
  • [WZH+13] Z. Wang, M. Zoghi, F. Hutter, D. Matheson, and N. De Freitas (2013) Bayesian optimization in high dimensions via random embeddings. In Twenty-Third International Joint Conference on Artificial Intelligence, Cited by: §1.

Appendix A Proofs of Section 3

Lemma 15.

If h has condition number Q, then for all xX, there is a ball of radius Q-1x-x* that is tangent at x and inside the sublevel set Lx(h).


Write h=gf such that f is α-strongly convex and β-smooth for some β=Qα and g is monotonically increasing. From the smoothness assumption, we have for any s,


Consider the ball B=(x-1βf(x),1βf(x)). For any yB, the above inequality implies f(y)f(x). Hence, when we apply g on both sides, we still have h(y)h(x) for all yB. Therefore, Bh(y).

By strong convexity, f(x)αx-x*. It follows that the radius of B is at least αβx-x*. ∎

Proof of Lemma 8.

Without loss of generality, consider the unit distance case where =1. Furthermore, it suffices to prove for the smallest possible radius r2=1-14n.

Since |r1-r2|r1+r2, the intersection B1B2 is composed of two hyperspherical caps glued end to end. We lower bound vol(B1B2) by the volume of the cap C1 of B1 that is contained in the intersection. Consider the triangle with sides r1,r2 and . From classic geometry, the height of C1 is

c1=12(1+r12-r22)>0. (1)

The volume of a spherical cap is [LI11],


where I is the regularized incomplete beta function defined as


where x[0,1] and a,b(0,). Note that for any fixed a and b, Ix(a,b) is increasing in x. Hence, in order to obtain a lower bound on vol(C1), we want to lower bound 1-c12r12 or equivalently, upper bound c12r12.

Write r1=α2n for some α[1,2]. From Eq. (1),




Since g(α)=8α+4α is convex in [1,2], g(α)max(g(1),g(2))=12. It follows that c1r1116n(12-1n)34n. So, 1-c12r121-916n. To complete the proof, note that Vn:=I1-916n(n+12,12) is increasing in n, and V1=14. As n goes to infinity, this value converges to 1 as B1B2. ∎

Proof of Lemma 7.

Let ν=15nQ. Let q=(1-ν)x+νx*. Let Bq=(cq,rq) be a ball that has q on its surface, lies inside q, and has radius rq=Q-1x-x*. Lemma 15 guarantees its existence.

Suppose that

vol(BxBq)14vol(Bx) (2)

and that a random sample y from Bx belongs to Bq, which happens with probability at least 14. Then, our guarantee follows by

f(y)-f(x*) f(q)-f(x*)

where the first line follows from Lemma 15 and second line from convexity of f.

Therefore, it now suffices to prove Eq. 2. To do so, we will apply Lemma 8 after showing that the radius of Bx and Bq are in the proper ranges. Let =x-cq and note that

x-q+rq (3)
(ν+Q-1(1-ν))x-x* (4)

Since x is outside of Bq, we also have

45Qbx-x*. (5)

It follows that


In the log2 space, our choice of k1 is equivalent to starting from log2C1 and sweeping through the range [log2C1,log2C2] at the interval of size 1. This is guaranteed to find a point between 2 and , which is also an interval of size 1. Therefore, there exists a k1 satisfying the theorem statement, and similarly, we can prove the existence of k2.

Finally, it remains to show that rq(1-1/(4n)). From Eq. (3), it suffices to show that x-q4n or equivalently νx-x*4n. From Eq. (4),


For any Q,n1, 1-ν45. So,

νQ(1-ν)-1=15n(1-ν)-114n (6)

and the proof is complete. ∎

Proof of Lemma 9.

Without loss of generality, let =1 and B2 is centered at the origin with radius r2 and B1 is centered at e1=(1,0,,0). Then, we simply want to show that


By Markov’s inequality, we see that i=2nXi22r12=2/n with probability at most 1/2. And since X1 is independent and r21-1/n, it suffices to show that


Since X1 has standard deviation at least r1/n1/(2n), we see that the probability of deviating at least a few standard deviation below is at least a constant. ∎

Proof of Theorem 10.

We can consider the projection of all points onto the column space of A and since the Gaussian sampling process is preserved, our proof follows from applying Theorem 7 restricted onto the k-dimensional subspace and using Lemma 9 in place of Lemma 8. ∎

Proof of Theorem 11.

By the boundedness of h, since f(x)-f(x*)60δkQg(𝐀), we see that g(𝐏𝐀x)-g(x*)60δkQg(𝐀)-2δ>0. By Lemma 9, we see that if we sample from a Gaussian distribution y𝒩(x,r2k𝐈), then if z* is the minimum of g(x) restricted to the column space of 𝐀, then


with constant probability. By boundedness on h, we know that h(y)h(x)+2δ. Furthermore, this also implies that g(𝐏𝐀x*)g(z*)+2δ. Therefore, we know that the decrease is at least

f(y)-f(x*) =g(𝐏𝐀y)-g(𝐏𝐀x*)+h(y)-h(x*)

Since f(x)-f(x*)10δkQg(A), we conclude that


and our proof is complete. ∎

Proof of Theorem 12.

Our main proof strategy is to show that progress can only be made with a radius size of O(log(nQ)/(nQ)); larger radii cannot find descent directions with high probability. Consider a simple ellipsoid function f(x)=xDx, where D is a diagonal matrix and D11D22Dnn, where WLOG we let D11=1 and Dii=Q for i>1. The optima is x*=0 with f(x*)=0.

Consider the region X={x=(x1,x2,,xn)|1x10.9,|xi|0.1/(Qn)}. Then, if we let vN(0,I) be a standard Gaussian vector, then for some radius r, we see that the probability of finding a descent direction is:

𝐏𝐫[f(x+rv)f(x)] =𝐏𝐫[(x1+rv1)2+i>1Dii(xi+rvi)2x12+i>1Diixi2]

By standard concentration bounds for sub-exponential variables, we have


Therefore, with exponentially high probability, i>1Xi2n/2. Also, since |xi|0.1/(Qn), Chernoff bounds give:


Therefore, with probability at least 1-1/(nQ)3, |i>1viXi|log(nQ)/Q.

If QrnΩ(log(nQ)), then we have

-Qi>1viXi-12Qri>1Xi2 -Ω(log(nQ))

We conclude that the probability of descent is upper bounded by 𝐏𝐫[v1X1-Ω(log(nQ))]. This probability is exactly Φ(-l), where Φ is the cumulative density of a standard normal and l=Ω(log(nQ)). By a naive upper bound, we see that

Φ(-l) =12πle-x2/2𝑑x

Since l=Ω(log(nQ)), we conclude that Therefore, we conclude that with probability at least 1-1/poly(nQ), we have f(y)-f(x*)f(x)-f(x*).

Otherwise, we are in the case that QrnO(log(nQ)). Arguing simiarly as before, with high probability, our objective function and each coordinate can change by at most O(log(nQ)/(Qn)).

Next, we extend our proof to any symmetric distribution 𝒟. Since 𝒟 is rotationally symmetric, if we parametrize v=(r,θ) is polar-coordinates, then the p.d.f. of any scaling of 𝒟 must take the form p(v)=pr(r)u(θ), where u(θ) induces the uniform distribution over the unit sphere. Therefore, if Y is a random variable that follows 𝒟, then we may write Y=Rv/v, where R is a random scalar with p.d.f pr(r) and v is a standard Gaussian vector and R,X are independent.

As previously argued, v[0.5n,1.5n] with exponentially high probability. Therefore, if RΩ(log(nQ)/Q), the same arguments will imply that Y is a descent direction with polynomially small probability. Thus, when Y is a descent direction, it must be that RΩ(log(nQ)/Q) and as argued previously, our lower bound follows similarly.

Appendix B Proofs of Section 4

Proof of Theorem 13.

By the Gaussian version of Theorem 7 (full rank version of Theorem 10), as long as our binary search sweeps between minimum search radius r35Qnx-x* and maximum search radius of the diameter of the whole space R=𝒳, the objective value will decrease multiplicatively by 1-15nQ in each iteration with constant probability. Therefore, if xt-x*2Qϵ and we set r=ϵn and R=𝒳, then with high probability, we expect f(xT)-f(x*)βQ2ϵ2 after T=O(nQlog(𝒳/(Qϵ))) iterations, where we note that F=f(x0)-f(x*)β𝒳2 by smoothness.

Otherwise, if there exists some xt such that xt-x*2Qϵ, then f(xT)-f(x*)f(xt)-f(x*)4βQ2ϵ2. Therefore, by strong convexity, we conclude that in either case, xT-x*2Q3/2ϵ. Finally note that each iteration uses a binary search that requires O(log(R/r))=O(log(n𝒳/ϵ)) function evaluations.

Therefore, by combining these bounds, we derive our result. The low-rank result follows from applying Theorem 10 and Theorem 11 instead. ∎

Proof of Theorem 14.

Let H=O(nQlog(Q)) be the number of iterations between successive radius halving and we initialize R=𝒳 and half R every H iterations. We call the iterations between two halving steps an epoch. We claim that xi-x0R for all iterations and proceed with induction on the epoch number. The base case is trivial.

Assume that xi-x0R for all iterations in the previous epoch and let iteration is be the start of the epoch and iteration is+H be the end of the epoch. Then, since xis-x*R, we see that f(xis)-f(x*)βR2 by smoothness. If R4Qxi-x*4QR for all i in the previous epoch, then by the Gaussian version of Theorem 7 (Theorem 10), since we do a binary sweep from R4Q to 4QR, we can choose 𝒟 accordingly so that we are guaranteed that our objective value will decrease multiplicatively by 1-15nQ with constant probability at a cost of O(log(Q)) function evaluations per iteration. This implies that with high probability, after O(nQlog(Q)) iterations, we conclude


Otherwise, there exists some 1jH such that xis+j-x*4QR or xis+j-x*R4Q. If it is the former, then by strong convexity, f(xis+j)-f(x*)αxis+j-x*22βR2, which contradicts the fact that f(xis)-f(x*)βR2 by smoothness. If it is the latter, then by smoothness, we reach the same conclusion:


Therefore, by strong convexity, we have


And our induction is complete. Therefore, we conclude that after log(𝒳/ϵ) epochs, we have xT-x*ϵ. Each epoch has H iterations, each with O(log(Q)) function evaluations and so our result follows.

The low-rank result follows from applying Theorem 10 and Theorem 11 instead. However, note that since we do not know the latent dimension k, we must extend the binary search to incur an extra log(n) factor in the binary search cost. ∎

Appendix C Figures

Figure 2: The average optimality gap by the condition number of the objective function.
Figure 3: The average optimality gap by the accuracy of the condition number estimate, where approximation factor is the ratio of estimated to true condition number. The dimension n=20.

C.1 BBOB Function Plots

Figure 4: Convergence plot for the BBOB Rastrigin Function.
Figure 5: Convergence plot for the BBOB BentCigar Function.
Figure 6: Convergence plot for the BBOB BuecheRastrigin Function.
Figure 7: Convergence plot for the BBOB DifferentPowers Function.
Figure 8: Convergence plot for the BBOB Discus Function.
Figure 9: Convergence plot for the BBOB Ellipsoidal Function.
Figure 10: Convergence plot for the BBOB Katsuura Function.
Figure 11: Convergence plot for the BBOB SchaffersF7 Function.
Figure 12: Convergence plot for the BBOB Ill-Conditioned SchaffersF7 Function.
Figure 13: Convergence plot for the BBOB SharpRidge Function.
Figure 14: Convergence plot for the BBOB Weierstass Function.

C.2 Mujoco Control Plots

Figure 15: Plot of maximum reward so far found by algorithm. Main line is the median trajectory across 3 runs.
Figure 16: Example of rewards found by all samples by algorithm.