L10b: Models of Computation

    An interesting question for computer architects is what capabilities must be included in the ISA? When we studied Boolean gates in Part 1 of the course, we were able to prove that NAND were universal, i.e., that we could implement any Boolean function using only circuits constructed from NAND gates.

    We can ask the corresponding question of our ISA: is it universal, i.e., can it be used to perform any computation? what problems can we solve with a von Neumann computer? Can the Beta solve any problem FSMs can solve? Are there problems FSMs can’t solve? If so, can the Beta solve those problems? Do the answers to these questions depend on the particular ISA?

    To provide some answers, we need a mathematical model of computation. Reasoning about the model, we should be able to prove what can be computed and what can’t. And hopefully we can ensure that the Beta ISA has the functionality needed to perform any computation.

    The roots of computer science stem from the evaluation of many alternative mathematical models of computation to determine the classes of computation each could represent. An elusive goal was to find a universal model, capable of representing *all* realizable computations. In other words if a computation could be described using some other well-formed model, we should also be able to describe the same computation using the universal model.

    One candidate model might be finite state machines (FSMs), which can be built using sequential logic. Using Boolean logic and state transition diagrams we can reason about how an FSM will operate on any given input, predicting the output with 100% certainty.

    Are FSMs the universal digital computing device? In other words, can we come up with FSM implementations that implement all computations that can be solved by any digital device?

    Despite their usefulness and flexibility, there are common problems that cannot be solved by any FSM. For example, can we build an FSM to determine if a string of parentheses (properly encoded into a binary sequence) is well-formed? A parenthesis string is well-formed if the parentheses balance, i.e., for every open parenthesis there is a matching close parenthesis later in the string. In the example shown here, the input string on the top is well-formed, but the input string on the bottom is not. After processing the input string, the FSM would output a 1 if the string is well-formed, 0 otherwise.

    Can this problem be solved using an FSM? No, it can’t. The difficulty is that the FSM uses its internal state to encode what it knows about the history of the inputs. In the paren checker, the FSM would need to count the number of unbalanced open parens seen so far, so it can determine if future input contains the required number of close parens. But in a finite state machine there are only a fixed number of states, so a particular FSM has a maximum count it can reach. If we feed the FSM an input with more open parens than it has the states to count, it won’t be able to check if the input string is well-formed.

    The “finite-ness” of FSMs limits their ability to solve problems that require unbounded counting. Hmm, what other models of computation might we consider? Mathematics to the rescue, in this case in the form of a British mathematician named Alan Turing.

    In the early 1930’s Alan Turing was one of many mathematicians studying the limits of proof and computation. He proposed a conceptual model consisting of an FSM combined with a infinite digital tape that could read and written under the control of the FSM. The inputs to some computation would be encoded as symbols on the tape, then the FSM would read the tape, changing its state as it performed the computation, then write the answer onto the tape and finally halting. Nowadays, this model is called a Turing Machine (TM). Turing Machines, like other models of the time, solved the “finite” problem of FSMs.

    So how does all this relate to computation? Assuming the non-blank input on the tape occupies a finite number of adjacent cells, it can be expressed as a large integer. Just construct a binary number using the bit encoding of the symbols from the tape, alternating between symbols to the left of the tape head and symbols to the right of the tape head. Eventually all the symbols will be incorporated into the (very large) integer representation.

    So both the input and output of the TM can be thought of as large integers, and the TM itself as implementing an integer function that maps input integers to output integers.

    The FSM brain of the Turing Machine can be characterized by its truth table. And we can systematically enumerate all the possible FSM truth tables, assigning an index to each truth table as it appears in the enumeration. Note that indices get very large very quickly since they essentially incorporate all the information in the truth table. Fortunately we have a very large supply of integers!

    We’ll use the index for a TM’s FSM to identify the TM as well. So we can talk about TM 347 running on input 51, producing the answer 42.

    There are many other models of computation, each of which describes a class of integer functions where a computation is performed on an integer input to produce an integer answer. Kleene, Post and Turing were all students of Alonzo Church at Princeton University in the mid-1930’s. They explored many other formulations for modeling computation: recursive functions, rule-based systems for string rewriting, and the lambda calculus. They were all particularly intrigued with proving the existence of problems unsolvable by realizable machines. Which, of course, meant characterizing the problems that could be solved by realizable machines.

    It turned out that each model was capable of computing exactly the same set of integer functions! This was proved by coming up with constructions that translated the steps in a computation between the various models. It was possible to show that if a computation could be described by one model, an equivalent description exists in the other model. This lead to a notion of computability that was independent of the computation scheme chosen. This notion is formalized by Church’s Thesis, which says that every discrete function computable by any realizable machine is computable by some Turing Machine. So if we say the function f(x) is computable, that’s equivalent to saying that there’s a TM that given x as an input on its tape will write f(x) as an output on the tape and halt.

    As yet there’s no proof of Church’s Thesis, but it’s universally accepted that it’s true. In general “computable” is taken to mean “computable by some TM”.

    If you’re curious about the existence of uncomputable functions, please see the optional video at the end of this lecture.

    Okay, we’ve decided that Turing Machines can model any realizable computation. In other words for every computation we want to perform, there’s a (different) Turing Machine that will do the job. But how does this help us design a general-purpose computer? Or are there some computations that will require a special-purpose machine no matter what?

    What we’d like to find is a universal function U: it would take two arguments, k and j, and then compute the result of running $T_k$ on input j. Is U computable, i.e., is there a universal Turing Machine $T_U$? If so, then instead of many ad-hoc TMs, we could just use $T_U$ to compute the results for any computable function.

    Surprise! U is computable and $T_U$ exists. If fact there are infinitely many universal TMs, some quite simple - the smallest known universal TM has 4 states uses 6 tape symbols. A universal machine is capable of performing any computation that can be performed by any TM!

    What’s going on here? k encodes a “program” - a description of some arbitrary TM that performs a particular computation. j encodes the input data on which to perform that computation. $T_U$ “interprets” the program, emulating the steps $T_k$ will take to process the input and write out the answer. The notion of interpreting a coded representation of a computation is a key idea and forms the basis for our stored program computer.

    The Universal Turing Machine is the paradigm for modern general-purpose computers. Given an ISA we want to know if it’s equivalent to a universal Turing Machine. If so, it can emulate every other TM and hence compute any computable function.

    How do we show our computer is Turing Universal? Simply demonstrate that it can emulate some known Universal Turing Machine. The finite memory on actual computers will mean we can only emulate UTM operations on inputs up to a certain size, but within this limitation we can show our computer can perform any computation that fits into memory.

    As it turns out this is not a high bar: so long as the ISA has conditional branches and some simple arithmetic, it will be Turing Universal.

    This notion of encoding a program in a way that allows it to be data to some other program is a key idea in computer science.

    We often translate a program Px written to run on some abstract high-level machine (eg, a program in C or Java) into, say, an assembly language program Py that can be interpreted by our CPU. This translation is called compilation.

    Much of software engineering is based on the idea of taking a program and using it as as component in some larger program.

    Given a strategy for compiling programs, that opens the door to designing new programming languages that let us express our desired computation using data structures and operations particularly suited to the task at hand.

    So what have learned from the mathematicians’ work on models of computation? Well, it’s nice to know that the computing engine we’re planning to build will be able to perform any computation that can be performed on any realizable machine. And the development of the universal Turing Machine model paved the way for modern stored-program computers. The bottom line: we’re good to go with the Beta ISA!

    We’ve discussed computable functions. Are there uncomputable functions?

    Yes, there are well-defined discrete functions that cannot be computed by any TM, i.e., no algorithm can compute f(x) for arbitrary finite x in a finite number of steps. It’s not that we don’t know the algorithm, we can actually prove that no algorithm exists. So the finite memory limitations of FSMs wasn’t the only barrier as to whether we can solve a problem.

    The most famous uncomputable function is the so-called Halting function. When TMs undertake a computation there two possible outcomes. Either the TM writes an answer onto the tape and halts, or the TM loops forever. The Halting function tells which outcome we’ll get: given two integer arguments k and j, the Halting function determines if the kth TM halts when given a tape containing j as the input.

    Let’s quickly sketch an argument as to why the Halting function is not computable. Well, suppose it was computable, then it would be equivalent to some TM, say $T_H$.

    So we can use $T_H$ to build another TM, $T_N$ (the “N” stands for nasty!) that processes its single argument and either LOOPs or HALTs. $T_N[X]$ is designed to loop if TM X given input X halts. And vice versa: $T_N[X]$ halts if TM X given input X loops. The idea is that $T_N[X]$ does the opposite of whatever $T_X[X]$ does. $T_N$ is easy to implement assuming that we have $T_H$ to answer the “halts or loops” question.

    Now consider what happens if we give N as the argument to $T_N$. From the definition of $T_N$, $T_N[N]$ will LOOP if the halting function tells us that $T_N[N]$ halts. And $T_N[N]$ will HALT if the halting function tells us that $T_N[N]$ loops. Obviously $T_N[N]$ can’t both LOOP and HALT at the same time! So if the Halting function is computable and $T_H$ exists, we arrive at this impossible behavior for $T_N[N]$. This tells us that $T_H$ cannot exist and hence that the Halting function is not computable.