next up previous
Next: Measuring Bias Through Optimal Up: OOPS on Universal Computers Previous: Basic Principles of OOPS

Essential Properties of OOPS

The following observations highlight important aspects of OOPS and clarify in which sense OOPS is optimal.

Observation 3.1   A program starting at $a_{last}$ and solving task $n$ will also solve all tasks up to $n$.

Proof (exploits the nature of self-delimiting programs): Obvious for $n=1$. For $n> 1$: By induction, the code between $a_{last}$ and $a_{frozen}$, which cannot be altered any more, already solves all tasks up to $n-1$. During its application to task $n$ it cannot request any additional tokens that could harm its performance on these previous tasks. So those of its prolongations that solve task $n$ will also solve tasks $1, \ldots, n-1$.

Observation 3.2   $a_{last}$ does not increase if task $n$ can be more quickly solved by testing prolongations of $q_{a_{last}:a_{frozen}}$ on task $n$, than by testing fresh programs starting above $a_{frozen}$ on all tasks up to $n$.

Observation 3.3   Once we have found an optimal universal solver for all tasks in the sequence, at most three quarters of the total future time will be wasted on searching for faster alternatives.

Ignoring hardware-specific overhead (e.g., for measuring time and switching between tasks), OOPS will lose at most a factor 2 through allocating half the search time to prolongations of $q_{a_{last}:a_{frozen}}$, and another factor 2 through the incremental doubling of time limits in LSEARCH (necessary because we do not know in advance the final time limit).

Observation 3.4   Given the initial bias and subsequent code bias shifts due to $q^1, q^2, \ldots, $ no bias-optimal searcher with the same initial bias will solve the current task set substantially faster than OOPS.

Bias shifts may greatly accelerate OOPS though. Consider a program $p$ solving the current task set, given current code bias $q_{0:a_{frozen}}$ and $a_{last}$. Denote $p$'s probability by $P(p)$ (for simplicity we omit the obvious conditions). The advantages of OOPS materialize when $P(p) >> P(p')$, where $p'$ is among the most probable fast solvers of the current task that do not use previously found code. Ideally, $p$ is already identical to the most recently frozen code. Alternatively, $p$ may be rather short and thus likely because it uses information conveyed by earlier found programs stored below $a_{frozen}$. For example, $p$ may call an earlier stored $q^i$ as a subprogram. Or maybe $p$ is a short and fast program that copies a large $q^i$ into state $s(r_j)$, then modifies the copy just a little bit to obtain $\bar{q}^i$, then successfully applies $\bar{q}^i$ to $r_j$. Clearly, if $p'$ is not many times faster than $p$, then OOPS will in general suffer from a much smaller constant slowdown factor than nonincremental LSEARCH, precisely reflecting the extent to which solutions to successive tasks do share useful mutual information, given the set of primitives for copy-editing them.

Observation 3.5   OOPS remains near-bias-optimal not only with respect to the initial bias but also with respect to any subsequent bias due to additional frozen code.

That is, unlike the learning rate-based bias shifts of ALS [65] (compare Section 2.4), those of OOPS do not reduce the probabilities of programs that were meaningful and executable before the addition of any new $q^i$. To see this, consider probabilities of fresh program prefixes examined by OOPS, including formerly interrupted program prefixes that once tried to access code for earlier solutions when there weren't any. OOPS cannot decrease such probabilities. But formerly interrupted prefixes may suddenly become prolongable and successful, once some solutions to earlier tasks have been stored. Thus the probabilities of such prolongations may increase from zero to a positive value, without any compensating loss of probabilities of alternative prefixes.

Therefore, unlike with ALS the acceleration potential of OOPS is not bought at the risk of an unknown slowdown due to nonoptimal changes of the underlying probability distribution (based on a heuristically chosen learning rate). As new tasks come along, OOPS remains near-bias-optimal with respect to any of its previous biases.

Observation 3.6   If the current task (say, task $n$) can already be solved by some previously frozen program $q^k$, then the probability of a solver for task $n$ is at least equal to the probability of the most probable program computing the start address $a(q^k)$ of $q^k$ and then setting instruction pointer $ip(n):=a(q^k)$.

Observation 3.7   As we solve more and more tasks, thus collecting and freezing more and more $q^i$, it will generally become harder and harder to identify and address and copy-edit useful code segments within the earlier solutions.

As a consequence we expect that much of the knowledge embodied by certain $q^j$ actually will be about how to access and copy-edit and otherwise use programs $q^i$ ($i<j$) previously stored below $q^j$.

Observation 3.8   Tested program prefixes may rewrite the probability distribution on their suffixes in computable ways (e.g., based on previously frozen $q^i$), thus temporarily redefining the probabilistic search space structure of LSEARCH, essentially rewriting the search procedure. If this type of metasearching for faster search algorithms is useful to accelerate the search for a solution to the current problem, then OOPS will automatically exploit this.

That is, while ALS [65] has a prewired way of changing probabilities, OOPS can learn such a way, as it searches among prefixes that may compute changes of their suffix probabilities in online fashion.

Since there is no fundamental difference between domain-specific problem-solving programs and programs that manipulate probability distributions and rewrite the search procedure itself, we collapse both learning and metalearning in the same time-optimal framework.

Observation 3.9   If the overall goal is just to solve one task after another, as opposed to finding a universal solver for all tasks, it suffices to test only on task $n$ in step 3.

For example, in an optimization context the $n$-th problem usually is not to find a solver for all tasks $1, \ldots, n$, but just to find an approximation to some unknown optimal solution such that the new approximation is better than the best found so far.

next up previous
Next: Measuring Bias Through Optimal Up: OOPS on Universal Computers Previous: Basic Principles of OOPS
Juergen Schmidhuber 2004-04-15

Back to OOPS main page