We present a novel, general, optimally fast, incremental way of searching for a universal algorithm that solves each task in a sequence of tasks. The Optimal Ordered Problem Solver (OOPS) continually organizes and exploits previously found solutions to earlier tasks, eciently searching not only the space of domain-specific algorithms, but also the space of search algorithms. Essentially we extend the principles of optimal nonincremental universal search to build an incremental universal learner that is able to improve itself through experience.

Solomonoff's optimal but noncomputable method for inductive inference assumes that observation sequences x are drawn from an recursive prior distribution p(x). Instead of using the unknown p() he predicts using the celebrated universal enumerable prior M() which for all exceeds any recursive p(), save for a constant factor independent of x. The simplicity measure M() naturally implements "Occam's razor " and is closely related to the Kolmogorov complexity of . However, M assigns high probability to certain data that are extremely hard to compute. This does not match our intuitive notion of simplicity. Here we suggest a more plausible measure derived from the fastest way of computing data. In absence of contrarian evidence, we assume that the physical world is generated by a computational process, and that any possibly infinite sequence of observations is therefore computable in the limit (this assumption is more radical and stronger than Solomonoff's).

The probability of observing $x_t$ at time $t$, given past observations $x_1...x_{t-1}$ can be computed with Bayes rule if the true generating distribution $\mu$ of the sequences $x_1x_2x_3...$ is known. If $\mu$ is unknown, but known to belong to a class $M$ one can base ones prediction on the Bayes mix $\xi$ defined as a weighted sum of distributions $ u\in M$. Various convergence results of the mixture posterior $\xi_t$ to the true posterior $\mu_t$ are presented. In particular a new (elementary) derivation of the convergence $\xi_t/\mu_t\to 1$ is provided, which additionally gives the rate of convergence. A general sequence predictor is allowed to choose an action $y_t$ based on $x_1...x_{t-1}$ and receives loss $\ell_{x_t y_t}$ if $x_t$ is the next symbol of the sequence. No assumptions are made on the structure of $\ell$ (apart from being bounded) and $M$. The Bayes-optimal prediction scheme $\Lambda_\xi$ based on mixture $\xi$ and the Bayes-optimal informed prediction scheme $\Lambda_\mu$ are defined and the total loss $L_\xi$ of $\Lambda_\xi$ is bounded in terms of the total loss $L_\mu$ of $\Lambda_\mu$. It is shown that $L_\xi$ is bounded for bounded $L_\mu$ and $L_\xi/L_\mu\to 1$ for $L_\mu\to \infty$. Convergence of the instantaneous losses is also proven.

The Bayesian framework is a well-studied and successful framework for inductive reasoning, which includes hypothesis testing and confirmation, parameter estimation, sequence prediction, classification, and regression. But standard statistical guidelines for choosing the model class and prior are not always available or can fail, in particular in complex situations. Solomonoff completed the Bayesian framework by providing a rigorous, unique, formal, and universal choice for the model class and the prior. I discuss in breadth how and in which sense universal (non-i.i.d.) sequence prediction solves various (philosophical) problems of traditional Bayesian sequence prediction. I show that Solomonoff’s model possesses many desirable properties: Strong total and future bounds, and weak instantaneous bounds, and in contrast to most classical continuous prior densities has no zero p(oste)rior problem, i.e. can confirm universal hypotheses, is reparametrization and regrouping invariant, and avoids the old-evidence and updating problem. It even performs well

An old dream of computer scientists is to build an optimally ecient universal problem solver. We show how to solve arbitrary computational problems in an optimal fashion inspired by Kurt Godel's celebrated self-referential formulas (1931). Our Godel machine's initial software includes an axiomatic description of: the Godel machine's hardware, the problem-speci c utility function (such as the expected future reward of a robot), known aspects of the environment, costs of actions and computations, and the initial software itself (this is possible without introducing circularity). It also includes a typically sub-optimal initial problem-solving policy and an asymptotically optimal proof searcher searching the space of computable proof techniques|that is, programs whose outputs are proofs. Unlike previous approaches, the self-referential Godel machine will rewrite any part of its software, including axioms and proof searcher, as soon as it has found a proof that this will improve its future performance, given its typically limited computational resources. We show that self-rewrites are globally optimal|no local minima!|since provably none of all the alternative rewrites and proofs (those that could be found by continuing the proof search) are worth waiting for.

Solomonoff completed the Bayesian framework by providing a rigorous, unique, formal, and universal choice for the model class and the prior. We discuss in breadth how and in which sense universal (non-i.i.d.) sequence prediction solves various (philosophical) problems of traditional Bayesian sequence prediction. We show that Solomonoff’s model possesses many desirable properties: Fast convergence and strong bounds, and in contrast to most classical continuous prior densities has no zero p(oste)rior problem, i.e. can confirm universal hypotheses, is reparametrization and regrouping invariant, and avoids the old-evidence and updating problem. It even performs well (actually better) in non-computable environments.

Solomonoff unified Occam's razor and Epicurus' principle of multiple explanations to one elegant, formal, universal theory of inductive inference, which initiated the field of algorithmic information theory. His central result is that the posterior of his universal semimeasure M converges rapidly to the true sequence generating posterior μ, if the latter is computable. Hence, M is eligible as a universal predictor in case of unknown μ. We investigate the existence and convergence of computable universal (semi)measures for a hierarchy of computability classes: finitely computable, estimable, enumerable, and approximable. For instance, M is known...

Most traditional artificial intelligence (AI) systems of the past 50 years are either very limited, or based on heuristics, or both. The new millennium, however, has brought substantial progress in the field of theoretically optimal and practically feasible algorithms for prediction, search, inductive inference based on Occam's razor, problem solving, decision making, and reinforcement learning in environments of a very general type. Since inductive inference is at the heart of all inductive sciences, some of the results are relevant not only for AI and computer science but also for physics, provoking nontraditional predictions based on Zuse's thesis of the computer-generated universe.

I postulate that human or other intelligent agents function or should function as follows. They store all sensory observations as they come—the data is ‘holy.’ At any time, given some agent’s current coding capabilities, part of the data is compressible by a short and hopefully fast program / description / explanation / world model. In the agent’s subjective eyes, such data is more regular and more beautiful than other data. It is well-known that knowledge of regularity and repeatability may improve the agent’s ability to plan actions leading to external rewards. In absence of such rewards, however, known beauty is boring. Then interestingness becomes the first derivative of subjective beauty: as the learning agent improves its compression algorithm, formerly apparently random data parts become subjectively more regular and beautiful. Such progress in data compression is measured and maximized by the curiosity drive: create action sequences that extend the observation history and yield previously unknown / unpredictable but quickly learnable algorithmic regularity. I discuss how all of the above can be naturally implemented

The Bayesian framework is ideally suited for induction problems. The probability of observing $x_t$ at time $t$, given past observations $x_1...x_{t-1}$ can be computed with Bayes' rule if the true generating distribution $\mu$ of the sequences $x_1x_2x_3...$ is known. The problem, however, is that in many cases one does not even have a reasonable guess of the true distribution. In order to overcome this problem a universal (or mixture) distribution $\xi$ is defined as a weighted sum or integral of distributions $ u\!\in\!\M$, where $\M$ is any countable or continuous set of distributions including $\mu$. This is a generalization of Solomonoff induction, in which $\M$ is the set of all enumerable semi-measures. It is shown for several performance measures that using the universal $\xi$ as a prior is nearly as good as using the unknown true distribution $\mu$. In a sense, this solves the problem of the unknown prior in a universal way. All results are obtained for general finite alphabet. Convergence of $\xi$ to $\mu$ in a conditional mean squared sense and of $\xi/\mu\to 1$ with $\mu$ probability $1$ is proven. The number of additional errors $E_\xi$ made by the optimal universal prediction scheme based on $\xi$ minus the number of errors $E_\mu$ of the optimal informed prediction scheme based on $\mu$ is proven to be bounded by $O(\sqrt{E_\mu})$. The prediction framework is generalized to arbitrary loss functions. A system is allowed to take an action $y_t$, given $x_1...x_{t-1}$ and receives loss $\ell_{x_t y_t}$ if $x_t$ is the next symbol of the sequence. No assumptions on $\ell$ are necessary, besides boundedness. Optimal universal $\Lambda_\xi$ and optimal informed $\Lambda_\mu$ prediction schemes are defined and the total loss of $\Lambda_\xi$ is bounded in terms of the total loss of $\Lambda_\mu$, similar to the error bounds. We show that the bounds are tight and that no other predictor can lead to smaller bounds. Furthermore, for various performance measures we show Pareto-optimality of $\xi$ in the sense that there is no other predictor which performs better or equal in all environments $ u\in\M$ and strictly better in at least one. So, optimal predictors can (w.r.t.\ to most performance measures in expectation) be based on the mixture $\xi$. Finally we give an Occam's razor argument that Solomonoff's choice $w_ u\sim 2^{-K( u)}$ for the weights is optimal, where $K( u)$ is the length of the shortest program describing $ u$. Furthermore, games of chance, defined as a sequence of bets, observations, and rewards are studied. The average profit achieved by the $\Lambda_\xi$ scheme rapidly converges to the best possible profit. The time needed to reach the winning zone is proportional to the relative entropy of $\mu$ and $\xi$. The prediction schemes presented here are compared to the weighted majority algorithm(s). Although the algorithms, the settings, and the proofs are quite different the bounds of both schemes have a very similar structure. Extensions to infinite alphabets, partial, delayed and probabilistic prediction, classification, and more active systems are briefly discussed.