Parity-SAT is the problem of determining whether a given CNF formula has an odd number of satisfying assignments. As a canonical ⊕P-complete problem, it represents a fundamental variant of the exact model counting problem (#SAT). Under the Strong Exponential Time Hypothesis (SETH), Parity-SAT admits no O^*((2-ε)ⁿ)-time or O^*((2-ε)^m)-time algorithm for any constant ε > 0, where n and m denote the numbers of variables and clauses, respectively. Thus, breaking the 2ⁿ or 2^m barrier appears impossible in full generality. In this work, we revisit this barrier through structural restrictions and a refined exploitation of parity. We study Parity-d-occ-SAT, where each variable appears in at most d clauses, and obtain three main results. First, we design {a randomized} O^*(2^{m(1-1/O(d))})-time algorithm, thereby breaking the 2^m barrier for every fixed d. Second, for the special case d = 2, we develop a significantly sharper branching algorithm running in O^*(1.1193ⁿ) time or O^*(1.3248^m) time. Third, leveraging the structural insights underlying the d = 2 case, we obtain an O^*(1.1052^L)-time algorithm for general Parity-SAT, where L denotes the formula length. All algorithms use only polynomial space. Notably, our running-time bounds are better than the best known bounds for the corresponding exact counting counterparts, highlighting a genuine algorithmic advantage of parity over counting. Conceptually, our results demonstrate that parity admits finer structural reductions and more efficient branching than exact model counting, and that bounded occurrence can be systematically leveraged to circumvent classical exponential barriers.

This paper studies the recursion-theoretic aspects of large-scale geometries of infinite strings, a subject initiated by Khoussainov and Takisaka (2017). We investigate several notions of quasi-isometric reductions between recursive infinite strings and prove various results on the equivalence classes of such reductions. The main result is the construction of two infinite recursive strings $α$ and $β$ such that $α$ is strictly quasi-isometrically reducible to $β$, but the reduction cannot be made recursive. This answers an open problem posed by Khoussainov and Takisaka.

This paper studies the complexity of operations on finite automata and the complexity of their decision problems when the alphabet is unary. Let $n$ denote the maximum of the number of states of the input finite automata considered in the corresponding results. The following main results are obtained: (1) Given two unary NFAs recognising $L$ and $H$, respectively, one can decide whether $L \subseteq H$ as well as whether $L = H$ in time $2^{O((n \log n)^{1/3})}$. The previous upper bound on time was $2^{O((n \log n)^{1/2})}$ as given by Chrobak (1986), and this bound was not significantly improved since then. (2) Given two unary UFAs (unambiguous finite automata) recognising $L$ and $H$, respectively, one can determine a UFA recognising $L \cup H$ and a UFA recognising complement of $L$, where these output UFAs have the number of states bounded by a quasipolynomial in $n$. However, in the worst case, a UFA for recognising concatenation of languages recognised by two $n$-state UFAs, uses $2^{Θ((n \log^2 n)^{1/3})}$ states. (3) Given a unary language $L$, if $L$ contains the word of length $k$, then let $L(k)=1$ else let $L(k)=0$. Let $ω_L$ be the $ω$-word $L(0)L(1)\ldots$ and let $\cal L$ be a fixed $ω$-regular language. The last section studies how difficult it is to decide, given an $n$-state UFA or NFA

It is shown that the parity game can be solved in quasi-polynomial time. The parameterized parity game---with $n$ nodes and $m$ distinct values (a.k.a. colors or priorities)---is proven to be in the class of fixed parameter tractable problems when parameterized over $m$. Both results improve known bounds, from runtime $n^{O(\sqrt{n})}$ to $O(n^{\log(m)+6})$ and from an XP algorithm with runtime $O(n^{\Theta(m)})$ for fixed parameter $m$ to a fixed parameter tractable algorithm with runtime $O(n^5+2^{m\log(m)+6m})$. As an application, it is proven that colored Muller games with $n$ nodes and $m$ colors can be decided in time $O((m^m \cdot n)^5)$; it is also shown that this bound cannot be improved to $2^{o(m \cdot \log(m))} \cdot n^{O(1)}$ in the case that the exponential time hypothesis is true. Further investigations deal with memoryless Muller games and multidimensional parity games.

X3SAT is the problem of whether one can satisfy a given set of clauses with up to three literals such that in every clause, exactly one literal is true and the others are false. A related question is to determine the maximal Hamming distance between two solutions of the instance. Dahllöf provided an algorithm for Maximum Hamming Distance XSAT, which is more complicated than the same problem for X3SAT, with a runtime of $O(1.8348^n)$; Fu, Zhou and Yin considered Maximum Hamming Distance for X3SAT and found for this problem an algorithm with runtime $O(1.6760^n)$. In this paper, we propose an algorithm in $O(1.3298^n)$ time to solve the Max Hamming Distance X3SAT problem; the algorithm actually counts for each $k$ the number of pairs of solutions which have Hamming Distance $k$.

This paper introduces and studies a notion of \emph{algorithmic randomness} for subgroups of rationals. Given a randomly generated additive subgroup $(G,+)$ of rationals, two main questions are addressed: first, what are the model-theoretic and recursion-theoretic properties of $(G,+)$; second, what learnability properties can one extract from $G$ and its subclass of finitely generated subgroups? For the first question, it is shown that the theory of $(G,+)$ coincides with that of the additive group of integers and is therefore decidable; furthermore, while the word problem for $G$ with respect to any generating sequence for $G$ is not even semi-decidable, one can build a generating sequence $β$ such that the word problem for $G$ with respect to $β$ is co-recursively enumerable (assuming that the set of generators of $G$ is limit-recursive). In regard to the second question, it is proven that there is a generating sequence $β$ for $G$ such that every non-trivial finitely generated subgroup of $G$ is recursively enumerable and the class of all such subgroups of $G$ is behaviourally correctly learnable, that is, every non-trivial finitely generated subgroup can be semantically identified in the limit (again assuming that the set of generators of $G$ is limit-recursive). On the other hand, the class of non-trivial finitely generated subgroups of $G$ cannot be syntactically identified in the limit with respect to any generating sequence for $G$. The present work thus contributes to a recent line of research studying algorithmically random infinite structures and uncovers an interesting connection between the arithmetical complexity of the set of generators of a randomly generated subgroup of rationals and the learnability of its finitely generated subgroups.

The present paper deals with the inductive inference of recursively enumerable languages from positive data (also called text). It introduces the learning models of \emph{verifiability} and \emph{comparability}. The input to a verifier is an index $e$ and a text of the target language $L$, and the learner has to \emph{verify} whether or not the index $e$ input is correct for the target language $L$. A comparator receives two indices of languages from the target class $\cL$ as input and has to decide in the limit whether or not these indices generate the same language. Furthermore, \emph{standardisability} is studied, where a \emph{standardiser} receives an index $j$ of some target language $L$ from the class $\cL$, and for every $L∈\cL$ there must be an index $e$ such that $e$ generates $L$ and the standardiser has to map every index $j$ for $L$ to $e$. Additionally, the common learning models of \emph{explanatory learning}, \emph{conservative explanatory learning}, and \emph{behaviourally correct learning} are considered. For almost all learning models mentioned above it is also appropriate to consider the number of times a learner changes its mind. In particular, if no mind change occurs then we obtain the \emph{finite} variant of the models considered. Occasionally, also learning with the help of an oracle is taken into consideration. The main goal of this paper is to figure out to what extent verifiability, comparability, and standardisability are helpful for the inductive inference of classes of recursively enumerable languages. Here we also distinguish between \emph{indexed families}, \emph{one-one enumerable classes}, and \emph{recursively enumerable classes}. Our results are manyfold, and an almost complete picture is obtained. In particular, for indexed families and recursively enumerable classes finite comparability, finite standardisability, and finite verifiability always imply finite learnability. If at least one mind change is allowed, then there are differences, i.e., for indexed families, comparability or verifiability imply conservative explanatory learning, but standardisability does not; still explanatory learning can be achieved.

We study the connections between the learnability of automatic families of languages and the types of text used to present them to a learner. More precisely, we study how restrictions on the number of times that a correct datum appears in a text influence what classes of languages are automatically learnable. We show that an automatic family of languages is automatically learnable from fat text iff it is automatically learnable from thick text iff it is verifiable from balanced text iff it satisfies Angluin's tell-tale condition. Furthermore, many automatic families are automatically learnable from exponential text. We also study the relationship between automatic learnability and verifiability and show that all automatic families are automatically partially verifiable from exponential text and automatically learnable from thick text.

Parity games play an important role in model checking and synthesis. In their paper, Calude et al. have recently shown that these games can be solved in quasi-polynomial time. We show that their algorithm can be implemented efficiently: we use their data structure as a progress measure, allowing for a backward implementation instead of a complete unravelling of the game. To achieve this, a number of changes have to be made to their techniques, where the main one is to add power to the antagonistic player that allows for determining her rational move without changing the outcome of the game. We provide a first implementation for a quasi-polynomial algorithm, test it on small examples, and provide a number of side results, including minor algorithmic improvements, a quasi bi-linear complexity in the number of states and edges for a fixed number of colours, and matching lower bounds for the algorithm of Calude et al.

It is shown that the parity game can be solved in quasipolynomial time. The parameterised parity game - with n nodes and m distinct values (aka colours or priorities) - is proven to be in the class of fixed parameter tractable (FPT) problems when parameterised over m. Both results improve known bounds, from runtime nO(√n) to O(nlog(m)+6) and from an XP-algorithm with runtime O(nΘ(m)) for fixed parameter m to an FPT-algorithm with runtime O(n5)+g(m), for some function g depending on m only. As an application it is proven that coloured Muller games with n nodes and m colours can be decided in time O((mm · n)5); it is also shown that this bound cannot be improved to O((2m · n)c), for any c, unless FPT = W[1].

The present work investigates inductive inference from the perspective of reverse mathematics. Reverse mathematics is a framework which relates the proof strength of theorems and axioms throughout many areas of mathematics in an interdisciplinary way. The present work looks at basic notions of learnability including Angluin's tell-tale condition and its variants for learning in the limit and for conservative learning. Furthermore, the more general criterion of partial learning is investigated. These notions are studied in the reverse mathematics context for uniformly and weakly represented families of languages. The results are stated in terms of axioms referring to domination and induction strength.

This paper investigates the complexity of verbal languages and pattern languages of Thurston automatic groups in terms of the Chomsky hierarchy. Here the language generated by a pattern is taken as the set of representatives of all strings obtained when chosing values for the various variables. For noncommutative free groups, it is shown that the complexity of the verbal and pattern languages (in terms of level on the Chomsky hierarchy) does not depend on the Thurston automatic representation and that verbal languages cannot be context-free (unless they are either the empty word or the full group). They can however be indexed languages. Furthermore, it is shown that in the general case, it might depend on the exactly chosen Thurston automatic representation which level a verbal language takes in the Chomsky hierarchy. There are examples of groups where, in an appropriate representation, all pattern languages are regular or context-free, respectively.

Within the frameworks of learning in the limit of indexed classes of recursive languages from positive data and automatic learning in the limit of indexed classes of regular languages (with automatically computable sets of indices), we study the problem of minimizing the maximum number of mind changes F_M(n) by a learner M on all languages with indices not exceeding n. For inductive inference of recursive languages, we establish two conditions under which F_M(n) can be made smaller than any recursive unbounded non-decreasing function. We also establish how F_M(n) is affected if at least one of these two conditions does not hold. In the case of automatic learning, some partial results addressing speeding up the function F_M(n) are obtained.

Abstract We investigate a new paradigm in the context of learning in the limit, namely, learningcorrection grammarsfor classes ofcomputably enumerable (c.e.)languages. Knowing a language may feature a representation of it in terms oftwogrammars. The second grammar is used to make corrections to the first grammar. Such a pair of grammars can be seen as a single description of (or grammar for) the language. We call such grammarscorrection grammars. Correction grammars capture the observable fact that peopledocorrect their linguistic utterances during their usual linguistic activities. We show that learning correction grammars for classes of c.e. languages in theTxtEx-mode(i.e., converging to a single correct correction grammar in the limit) is sometimes more powerful than learning ordinary grammars even in theTxtBc-model (where the learner is allowed to converge to infinitely many syntactically distinct but correct conjectures in the limit). For eachn≥ 0. there is a similar learning advantage, again in learning correction grammars for classes of c.e. languages, but where we compare learning correction grammars that maken+ 1 corrections to those that makencorrections. The concept of a correction grammar can be extended into the constructive transfinite, using the idea of counting-down from notations for transfinite constructive ordinals. This transfinite extension can also be conceptualized as being about learning Ershov-descriptions for c.e. languages. Forua notation in Kleene's general system (O, <o) of ordinal notations for constructive ordinals, we introduce the concept of anu-correction grammar, whereuis used to bound the number of corrections that the grammar is allowed to make. We prove a general hierarchy result: ifuandvare notations for constructive ordinals such thatu

For the natural notion of splitting classes into two disjoint subclasses via a recursive classifier working on texts, the question of how these splittings can look in the case of learnable classes is addressed. Here the strength of the classes is compared using the strong and weak reducibility from intrinsic complexity. It is shown that, for explanatorily learnable classes, the complete classes are also mitotic with respect to weak and strong reducibility, respectively. But there is a weakly complete class that cannot be split into two classes which are of the same complexity with respect to strong reducibility. It is shown that, for complete classes for behaviorally correct learning, one-half of each splitting is complete for this learning notion as well. Furthermore, it is shown that explanatorily learnable and recursively enumerable classes always have a splitting into two incomparable classes; this gives an inductive inference counterpart of the Sacks splitting theorem from recursion theory.

Abstract Freivalds defined an acceptable programming system independent criterion for learning programs for functions in which the final programs were required to be both correct and “nearly” minimal size. i.e.. within a computable function of being purely minimal size. Kinber showed that this parsimony requirement on final programs limits learning power. However, in scientific inference, parsimony is considered highly desirable. Alim-computable functionis (by definition) one calculable by a total procedure allowed to change its mind finitely many times about its output. Investigated is the possibility of assuaging somewhat the limitation on learning power resulting from requiring parsimonious final programs by use of criteria which require the final, correct programs to be “not-so-nearly” minimal size, e.g., to be within a lim-computable function of actual minimal size. It is shown that some parsimony in the final program is thereby retained, yet learning power strictly increases. Considered, then, are lim-computable functions as above but for whichnotations forconstructive ordinals are used to bound the number of mind changes allowed regarding the output. This is a variant of an idea introduced by Freivalds and Smith. For this ordinal notation complexity bounded version of lim-computability, the power of the resultant learning criteria form finely graded, infinitely ramifying, infinite hierarchies intermediate between the computable and the lim-computable cases. Some of these hierarchies, for the natural notations determining them, are shown to be optimally tight.

Smith, Wiehagen and Zeugmann (1997) showed an interesting connection between learning with bounded number of mind changes from informants and classification from informant. They showed that if an indexed family of languages L is learnable via informants, using at most m mind changes, then one can partition 2N, the class of all languages, into m + 2 subclasses L1,..., Lm+2 such that (1) � i∈{1,2,...,m+1} Li = L, and (2) (L1,..., Lm+2) can be classified from informants. However Smith, Wiehagen and Zeugmann (1997) left open whether a similar result also holds for learning from texts. We show that such a result does not hold for texts.

Abstract Acceptable programming systems have many nice properties like s-m-n-Theorem, Composition and Kleene Recursion Theorem. Those properties are sometimes called control structures, to emphasize that they yield tools to implement programs in programming systems. It has been studied, among others by Riccardi and Royer, how these control structures influence or even characterize the notion of acceptable programming system. The following is an investigation, how these control structures behave in the more general setting of complete numberings as defined by Mal'cev and Eršov.

Proceedings of the Annual ACM Conference on Computational Learning Theory

Article Free Access Share on Robust learning aided by context Authors: John Case Department of CIS, University of Delaware, Newark, DE Department of CIS, University of Delaware, Newark, DEView Profile , Sanjay Jain Department of Information Systems and Computer Science, National University of Singapore, Singapore 119260, Republic of Singapore Department of Information Systems and Computer Science, National University of Singapore, Singapore 119260, Republic of SingaporeView Profile , Matthias Ott Institut für Logik, Komplexität und Deduktionssysteme Universität Karlsruhe, 76128 Karlsruhe, Germany Institut für Logik, Komplexität und Deduktionssysteme Universität Karlsruhe, 76128 Karlsruhe, GermanyView Profile , Arun Sharma School of Computer Science and Engineering, University of New South Wales, Sydney 2052, Australia School of Computer Science and Engineering, University of New South Wales, Sydney 2052, AustraliaView Profile , Frank Stephan Mathematisches Institut, Universität Heidelberg, 69120 Heidelberg, Germany Mathematisches Institut, Universität Heidelberg, 69120 Heidelberg, GermanyView Profile Authors Info & Claims COLT' 98: Proceedings of the eleventh annual conference on Computational learning theoryJuly 1998 Pages 44–55https://doi.org/10.1145/279943.279952Published:24 July 1998Publication History 1citation230DownloadsMetricsTotal Citations1Total Downloads230Last 12 Months8Last 6 weeks0 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteeReaderPDF

Proceedings of the Annual ACM Conference on Computational Learning Theory

. In an earlier paper, Kinber and Stephan posed an open problem about whether every class of languages, which can be identified strong monotonically, can also be identified by a set-driven machine. This question is solved in this paper. The answer to the question depends on whether the machines are required to be total or not! The solution of this result uncovers a finer gradation of the notion of setdrivenness.

Abstract Limiting identification of r.e. indexes for r.e. languages (from a presentation of elements of the language) and limiting identification of programs for computable functions (from a graph of the function) have served as models for investigating the boundaries of learnability. Recently, a new approach to the study of “intrinsic” complexity of identification in the limit has been proposed. This approach, instead of dealing with the resource requirements of the learning algorithm, uses the notion of reducibility from recursion theory to compare and to capture the intuitive difficulty of learning various classes of concepts. Freivalds, Kinber, and Smith have studied this approach for function identification and Jain and Sharma have studied it for language identification. The present paper explores the structure of these reducibilities in the context of language identification. It is shown that there is an infinite hierarchy of language classes that represent learning problems of increasing difficulty. It is also shown that the language classes in this hierarchy are incomparable, under the reductions introduced, to the collection of pattern languages. Richness of the structure of intrinsic complexity is demonstrated by proving that any finite, acyclic, directed graph can be embedded in the reducibility structure. However, it is also established that this structure is not dense. The question of embedding any infinite, acyclic, directed graph is open.

Most theoretical models of inductive inference make the idealized assumption that the data available to a learner is from a single and accurate source. The subject of inaccuracies in data emanating from a single source has been addressed by several authors. The present paper argues in favor of a more realistic learning model in which data emanates from multiple sources, some or all of which may be inaccurate. Three kinds of inaccuracies are considered: spurious data (modeled as noisy texts), missing data (modeled as incomplete texts), and a mixture of spurious and missing data (modeled as imperfect texts). Motivated by the above argument, the present paper introduces and theoretically analyzes a number of inference criteria in which a learning machine is fed data from multiple sources, some of which may be infected with inaccuracies. The learning situation modeled is the identification in the limit of programs from graphs of computable functions. The main parameters of the investigation are: the kind of inaccuracy, the total number of data sources, the number of faulty data sources which produce data within an acceptable bound, and the bound on the number of errors allowed in the final hypothesis learned by the machine. Sufficient conditions are determined under which, for the same kind of inaccuracy, for the same bound on the number of errors in the final hypothesis, and for the same bound on the number of inaccuracies, learning from multiple texts, some of which may be inaccurate, is equivalent to learning from a single inaccurate text. The general problem of determining when learning from multiple inaccurate texts is a restriction over learning from a single inaccurate text turns out to be combinatorially very complex. Significant partial results are provided for this problem. Several results are also provided about conditions under which the detrimental effects of multiple texts can be overcome by either allowing more errors in the final hypothesis or by reducing the number of inaccuracies in the texts. It is also shown that the usual hierarchies resulting from allowing extra errors in the final program (results in increased learning power) and allowing extra inaccuracies in the texts (results in decreased learning power) hold. Finally, it is demonstrated that in the context of learning from multiple inaccurate texts, spurious data is better than missing data, which in turn is better than a mixture of spurious and missing data.

This paper provides positive and negative results on algorithmically synthesizing, from grammars and from decision procedures for classes of languages, learning machines for identifying, from positive data, grammars for the languages in those classes. In the process, the uniformly decidable classes of recursive languages that can be behaviorally correctly identified from positive data are surprisingly characterized by Angluin's 1980 Condition 2 (the subset principle for preventing overgeneralization) . 1 Introduction In the context of learning programs in the limit for functions [KW80, AS83, OSW86b], a variety of enumeration techniques [Gol67, BB75, Wie90, Ful90b] are important and ubiquitous. Here is the archetypal case. Suppose one has an r.e. set P of programs for computing total functions. One proceeds as follows. At each point that one is given data about an input function, one conjectures the first program, if any, listed in P that is correct on the data seen through that point....

Article Free Access Share on Elementary formal systems, intrinsic complexity, and procrastination Authors: Sanjay Jain Department of ISCS, National University of Singapore, Singapore 119260, Republic of Singapore Department of ISCS, National University of Singapore, Singapore 119260, Republic of SingaporeView Profile , Arun Sharma School of Computer Science and Engineering, The University of New South Wales, Sydney, NSW 2052, Australia School of Computer Science and Engineering, The University of New South Wales, Sydney, NSW 2052, AustraliaView Profile Authors Info & Claims COLT '96: Proceedings of the ninth annual conference on Computational learning theoryJanuary 1996 Pages 181–192https://doi.org/10.1145/238061.238093Published:01 January 1996Publication History 6citation230DownloadsMetricsTotal Citations6Total Downloads230Last 12 Months7Last 6 weeks4 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteeReaderPDF

Consider a scenario in which an algorithmic machine, M, is being fed the graph of a computable function ƒ. M is said to finitely identify ƒ just in case, after inspecting a finite portion of the graph of ƒ, it emits its first conjecture, which is a program for ƒ, and it never abandons this conjecture thereafter. A team of machines is a multiset of such machines. A team is said to be successful just in case each member of some nonempty subset, of predetermined size, of the team is successful, The ratio of the number of machines required to be successful to the size of the team is referred to as the success ratio of the team. The present paper investigates the finite identification of computable functions by teams of learning machines. The results presented complete the picture for teams with success ratio 12 and greater. It is shown that at success ratio 12, introducing redundancy in the team can result in increased learning power. In particular, it is established that larger collections of functions can be learned by employing teams of 4 machines and requiring at least 2 to be successful than by employing teams of 2 machines and requiring at least 1 to be successful. Surprisingly, it is also shown that introducing further redundancy at success ratio 12 does not yield any extra learning power. In particular, it is shown that the collections of functions that can be finitely identified by a team of 2m machines requiring at least m to be successful is the same as: the collections of functions that can be finitely identified by a team of 4 machines requiring at least 2 to be successful, if is even, and the collections of functions that can be identified by a team of 2 machines requiring at least 1 to be successful, if is odd.

A new investigation of the complexity of language identification is undertaken using the notion of reduction from recursion theory and complexity theory. The approach, referred to as the intrinsic complexity of language identification, employs notions of “weak” and “strong” reduction between learnable classes of languages. The intrinsic complexity of several classes are considered and the results agree with the intuitive difficulty of learning these classes. Several complete classes are shown for both the reductions and it is also established that the weak and strong reductions are distinct.

A new identification criterion, motivated by notions of successively improving approximations in the philosophy of science, is defined. It shown that the class of recursive functions is identifiable under this criterion. This result is extended to apply to somewhat more realistic types of data than usual. This criterion is then modified to consider restrictions on the quality of approximations, and the new criteria are compared to existing criteria.

Abstract A generator program for a computable function (by definition) generates an infinite sequence of programs all but finitely many of which compute that function. Machine learning of generator programs for computable functions is studied. To motivate these studies partially, it is shown that, in some cases, interesting global properties for computable functions can be proved from suitable generator programs which cannot be proved from any ordinary programs for them. The power (for variants of various learning criteria from the literature) of learning generator programs is compared with the power of learning ordinary programs. The learning power in these cases is also compared to that of learning limiting programs , i.e., programs allowed finitely many mind changes about their correct outputs.

192-198

Machine learning of limit programs (i.e., programs allowed finitely many mind changes about their legitimate outputs) for computable functions is studied. Learning of iterated limit programs is also studied. To partially motivate these studies, it is shown that, in some cases, interesting global properties of computable functions can be proved from suitable (n+1)-iterated limit programs for them which can not be proved from any n-iterated limit programs for them. It is shown that learning power is increased when (n+1)-iterated limit programs rather than n-iterated limit programs are to be learned. Many tradeoff results are obtained regarding learning power, number (possibly zero) of limits taken, program size constraints, and number of errors tolerated in final programs learned.

Suppose LC 1 and LC 2 are two machine learning classes each based on a criterion of success. Suppose, for every machine which learns a class of functions according to the LC 2 criterion of success, there is a machine which learns this class according to the LC 2 criterion. In the case where the converse does not hold LC, is said to be separated from LC 2. It is shown that for many such separated learning classes from the literature a much stronger separation holds: (∀𝒞∈LC 1) (∃𝒞' ∈LC 2 - LC 1(( [' ⊃𝒞] It is also shown that there is a pair of separated learning classes from the literature for which the stronger separation above does not hold. A philosophical heuristic toward the design of artificially intelligent learning programs is presented with each strong separation result.

The effect of a partial explanation as additional information in the learning process is investigated. A scientist performs experiments to gather experimental data about some phenomenon, and then tries to construct an explanation (or theory) for the phenomenon. A plausible model for the practice of science is an inductive inference machine (scientist) learning a program (explanation) from a graph (set of experiments) of a recursive function (phenomenon). It is argued that this model of science is not an adequate one, as scientists, in addition to performing experiments, make use of some approximate partial explanation based on the “state of the art” knowledge about that phenomenon. An attempt has been made to model this partial explanation as additional information in the scientific process. It is shown that the inference capability of machines is improved in the presence of such a partial explanation. The quality of this additional information is modeled using certain “density” notions. It is shown that additional information about a “better” quality partial explanation enhances the inference capability of learning machines as scientists more than a “not so good” partial explanation. Similar enhancements to inference of approximations, a more sophisticated model of science, are demonstrated. Inadequacies in Gold's paradigm of language learning are investigated. It is argued that Gold's model fails to incorporate certain additional information that children get from their environment. Children are sometimes told about some grammatical rule that enumerates elements of the language. It is argued that these rules are a kind of additional information. They enable children to see in advance elements that are yet to appear in their environments. Also, children are being given some information about what is not in the language. Sometimes, they are rebuked for making incorrect utterances, or are told of a rule that enumerates certain non-elements of the language. An attempt has been made to extend Gold's model to incorporate both the above types of additional information. It is shown that either type of additional information enhances the learning capability of formal language learning devices.