The fox knows many things; the hedgehog knows one big thing.

Digital Foxes

The successful fox must know more than the sum of what the hedgehogs know, for it must know the connections from one thing to another. This fact is key to the design of computer systems for solving certain kinds of problems.

I keep coming across domains with the following structure:

  1. One wishes to solve a large, complex problem that fits into some uniform formalism (for example, a detailed simulation of a car as a system of 10,000 ODEs).

  2. The formalism has a variety of methods developed for solving problems stated in that formalism (e.g., various kinds of integrators). These methods are hedgehogs in that they treat the entire problem uniformly, possibly with state that’s global to the whole solution (e.g., the step size).

  3. The problem, being large, has many different parts, with different solution characteristics (e.g., some variables change slowly and smoothly, others oscillate rapidly).

  4. The performance of each hedgehog method on the whole problem is determined by its performance on the worst part of the problem (e.g., the step size has to be small enough to trace the highest-frequency oscillation in the model).

  5. Different methods (and different settings of the global state) have different weaknesses (e.g., high-order integrators are good for very smooth systems, because they can achieve large step sizes).

It is therefore desirable to solve such large problems like a fox: partition the large problem into smaller subproblems, that are more uniform in their solution characteristics, and solve each part with the most appropriate method.

On the one hand, such foxy partitioning is eased because these problems tend to be sparse, and have natural cliques, where variables within a clique interact with each other much more strongly and densely than with variables in a different clique.

On the other hand, partition is difficult, because the parts do interact, so the solution methods need to communicate partial or approximate solutions to each other during the solving process. Cross-method communication is especially difficult because each method is developed with its own invariants about what it communicates to itself during its progress and how, and, being a hedgehog, does not understand the communications of other hedgehog methods. Translation is therefore necessary.

This pattern seems to appear quite often at the cutting edge of computational science. Some formalisms where this phenomenon is encountered, with example problems:

  • Systems of ordinary differential equations or differential algebraic equations (the car example, with parts of greatly differing stiffness).

  • Coupled partial differential equations (e.g., detailed simulation of water+temperature flow through a system of pipes, pools, and tanks: those of the pipes that are narrow enough can be modeled with 1D PDEs, those of the pools that are shallow enough can be modeled with 2D PDEs, and the tanks with 3D PDEs).

  • Model inference in machine learning (different model shapes, like Gaussian mixtures and HMMs, can easily appear in the same compound model).

  • This is exacerbated by the rise of probabilistic programming languages, which easily generate models with long chains of deterministically-linked variables—the only place where exhaustive search shines as an inference method.

  • Constraint satisfaction

One common theoretical thread is that all these problem domains are at least NP-hard. This means that such problems are unsolvable in the general case (unless P = NP), which is why every hedgehog has weaknesses. Of course, this also means that any fox—any combination of methods, or any system for selecting methods—must also fail in some cases (unless P = NP). However, many particular instances of interest are solvable; and the cutting edge occurs where they are solvable only by compound methods, not by uniform application of basic ones.

It is therefore desirable to develop foxy computer systems: ones that allow easy (and, where possible, automatic) partitioning of a problem into appropriate parts, selection of methods for solving the parts, and composition thereof into a compound method for solving the whole problem.

How to Breed a Digital Fox

How might such a digital fox be developed? It seems to me that a system like this must have the following pieces:

  • Problem language: A language for specifying instances of the problem class (e.g., the Modelica language for DAEs, various probabilistic programming languages).

  • This language must of necessity have means of abstraction, in order for human authors of problem instances to be able to handle the complexity of the problems they specify (e.g., one does not write out 10,000 ODEs all in one big table). Are these abstraction boundaries always going to be the same as the desirable partitions of the problem into parts? If not, how should the partitions be specified or hinted at?

  • Hedgehogs: A library of primitive solution methods (e.g., 4th order Runge-Kutta, Gibbs sampling). I do not know to what extent such libraries already exist as reusable software in each domain. However, the definition of “a method” is that practitioners, after suitable study, know how to implement an instance of that method specialized to any particular problem they wish to solve with it. The trick here is to extract that knowledge into software definitions that specify enough of the method to automatically construct efficient specializations, but not so much as to over-restrict applicability.

  • Bridges: A library of bridges for communicating between solution methods. In my mind, this is the key missing piece in all proposed foxes I have seen so far—more below. In particular, I have seen no evidence of the knowledge needed to make this library appearing in the literature of any field, and I do not know for which domains this knowledge exists even in the minds of practitioners.

  • Plan language: A language for specifying problem-specific compound methods. A specification of a problem-specific compound method is a decomposition of the problem into parts, a specification of what methods are to be used to solve each part, and a specification of the bridges for communicating between them. Compound methods can have recursive structure in the sense that the method for solving a part of the problem can itself be compound. This part is a little tricky too, because one wants to be flexible about dynamic problem structure (e.g., Chinese Restaurant Process introducing more variables at solution runtime, or conditional triggers in an ODE simulation changing the applicable equations).

  • Plan executor: A mechanism for efficiently executing problem-specific compound solution methods. I presume this is going to be some combination of query planner, compiler, and runtime system, possibly using profiling information, etc. The better this mechanism is, the easier becomes the task of writing the libraries of primitive methods and bridges (the more the system can figure out itself, the less needs to be specified), and the less cumbersome becomes the language of compound methods.

  • Planner: A subsystem for evaluating, searching, suggesting, or inferring compound methods automatically based on the specification of the problem.

While it is important that these six pieces are different, and responsible for different jobs, I’m afraid that, with the notable exception of the planner, they must all be invented essentially together. What the hedgehogs are is presumably known at the start, but how they should be specified or implemented depends critically on the executor and the bridges. The bridges obviously depend critically on the hedgehogs they are bridging. The design of the plan language depends upon the composition constraints imposed by the hedgehogs and the qualities of the bridges, as well as how much it needs to tell the executor. The design of the plan language also back-influences the hedgehogs and bridges. The problem language is intimately tied to the plan language because they both talk about different perspectives on the same thing. And conversely, the executor depends upon what it will be executing. And all five of these essential pieces are complicated, so should be designed iteratively. This is exactly the kind of tight ball of design dependencies that should be neither rushed nor partitioned across organizations, but should rest with a close-knit team that has enough time to iterate on all the pieces together until a good solution emerges.

You will note that I omitted the planner from the previous admonition. I view the planner as optional because a digital fox without a planner would already be very useful: rather than programming solutions to their problems in C++, practitioners could program them in the plan language; even without any further assistance, I expect the gains in productivity to be tremendous.

In fact, I think the automatic method invention problem is something of a red herring. It is attractive to people who propose to build foxes because it looks like the kind of problem the fox is meant to solve, so they are tempted to apply their favorite methods from the discipline to try solving it. However, the method invention problem is guaranteed to be intractable in general (unless P = NP) because the underlying problems are intractable in general. What’s worse, the hope of automating this step can lead to an obscure and hard-to-use design for the plan language; which will turn the overall system into a usability nightmare in the cases where the planner invents a bad plan. I am therefore of the opinion that effort is better directed at getting the other parts right; and a clear design for the rest would enable a planner to be retrofit over the working fox later, after the nature of the method invention problem in this domain is better understood. Yes, doing so will probably require modifying the executor to instrument execution and give feedback about the quality of the plan, but that does not seem too steep a price to pay for not having to think too hard about the planner when working out the other five parts.

By the way, I see making a digital fox as a form of defeat. In some sense, all the understanding of the domain is contained in the available hedgehogs; and combining them into a fox introduces enormous additional complexity for no “fundamental” gain in coverage.1 This is not only complexity of the implementation, which is bad enough, but also complexity of the interface: a fox can fail in any of the ways its component methods can fail, and in additional ways introduced by method selection and method combination. So it’s much better to solve your problem with a better hedgehog if you can—making a fox is a last resort. But new hedgehogs are hard to invent, and the time comes when a field gives up and starts thinking about foxes. On the bright side, a high-quality fox should be able to capture new hedgehogs as they emerge from their dens.

On Bridges

The thing that makes some procedure a (hedgehog) solution method is that it sets up and uses some invariants for representing partial or approximate solutions to the problem, and iteratively expands or improves them. In general, for any two different hedgehogs, the invariants, and even the information communicated by them, will be different. In order to stitch more than one method together, then, it will be necessary to adjust the information one method generates for use by the other; possibly both ways if the flow of information in the solution is bidirectional.

This problem has an easy version and a hard version. The easy version occurs when two methods really need the same information to flow across the link, and that information just happens to need some kind of format conversion because of accidents of internal representations in the software implementing the methods. This version of the problem can probably be eliminated by implementing all the methods on a common substrate with common choices for representation of the same kinds of information. The hard version is when the methods actually need different information: for instance, an inference method that deals in marginal distributions uses different information from one that deals in samples. However, since both are ultimately about probability, there is hope that it may be possible to make an explicit bridge to translate, perhaps approximately, the information produced by one method into the information consumed by the other.

I can envision two kinds of bridges: ones that operate across a link in the structure of information flow, communicating across the link to different methods; and ones that operate at a node, reconciling overlapping information of different kinds. In the ODE domain, a link-type bridge would be a way to use an equation some of whose variables were being integrated exclusively by one integrator and some by another to exchange appropriate information between them. A node-type bridge, in contrast, would occur if we allowed the sets of variables integrated by different methods to overlap, and would be a way to reconcile the different values produced by the different methods for the same variable, and communicate information between the integrators through the results of reconciliation. I suppose hybrid bridges that incorporated aspects of the link-type and the node-type might also be possible.

My understanding is that bridge-building is not currently well understood. I hear rumors that people who carry out large simulations (ODEs, PDEs, etc) tend to implement such bridges, apparently in an ad-hoc way for each problem. But at least the knowledge is there of how to do it in principle. I don’t know whether such knowledge even exists for machine learning problems—has anyone successfully attacked a compound ML model using, say, MCMC to infer one part and a variational method to infer another?

Extant Foxes

When I started writing this essay, I didn’t think I knew of any digital foxes, but in retrospect the good old relational database management system fits the bill: SQL, together with the data schema, indexes, and table statistics, form the problem language; there are many different kinds of table scans, index searches, and result data structures available for performing parts of a query; the query plan language (which you can read with the SQL EXPLAIN statement) specifies problem-specific compound methods; and the query planner is the optional part that automatically searches plan space for good candidates.

The operating system kernel seems to stretch the analogy a little, but it does have the essential feature of managing multiple different hedgehogs (processes), intermediating communication between them (signals, pipes, sockets, files), and adjudicating use of resources (CPU, screen, keyboard, network card). There is less of a definite problem to solve, though, and method selection is pretty manual, so I don’t know whether Linux is a fox in this sense.

For the mathematically inclined, a sheaf is the only example I know about of a mathematical structure which is a fox. The gluing axiom is analogous to the all-important bridges.

Notes


  1. Actually, my main point, to the extent I have one, is that this is not so. The fox must know how to connect the hedgehogs together and make them cooperate; which perhaps necessitates a still-deeper understanding of the domain than each hedgehog embodies. And, of course, the whole point of a fox is that it does lead to better coverage, by covering problems that have parts that stymie each hedgehog.