arXiv subjects

Andras Lorincz

Publications and source records attributed to Andras Lorincz.

32 records · Page 2Linked to original sources

arXiv2007

Reinforcement Learning with Linear Function Approximation and LQ control Converges

Reinforcement learning is commonly used with function approximation. However, very few positive results are known about the convergence of function approximation based RL control algorithms. In this paper we show that TD(0) and Sarsa(0) with linear function approximation is convergent for a simple class of problems, where the system is linear and the costs are quadratic (the LQ control problem). Furthermore, we show that for systems with Gaussian noise and non-completely observable states (the LQG problem), the mentioned RL algorithms are still convergent, if they are combined with Kalman filtering.

cs.LG↗

arXiv2006

Low-complexity modular policies: learning to play Pac-Man and a new framework beyond MDPs

In this paper we propose a method that learns to play Pac-Man. We define a set of high-level observation and action modules. Actions are temporally extended, and multiple action modules may be in effect concurrently. A decision of the agent is represented as a rule-based policy. For learning, we apply the cross-entropy method, a recent global optimization algorithm. The learned policies reached better score than the hand-crafted policy, and neared the score of average human players. We argue that learning is successful mainly because (i) the policy space includes the combination of individual actions and thus it is sufficiently rich, (ii) the search is biased towards low-complexity policies and low complexity solutions can be found quickly if they exist. Based on these principles, we formulate a new theoretical framework, which can be found in the Appendix as supporting material.

cs.LG↗

arXiv2006

Real and Complex Independent Subspace Analysis by Generalized Variance

Here, we address the problem of Independent Subspace Analysis (ISA). We develop a technique that (i) builds upon joint decorrelation for a set of functions, (ii) can be related to kernel based techniques, (iii) can be interpreted as a self-adjusting, self-grouping neural network solution, (iv) can be used both for real and for complex problems, and (v) can be a first step towards large scale problems. Our numerical examples extend to a few 100 dimensional ISA tasks.

math.ST↗

arXiv2006

Separation Theorem for K-Independent Subspace Analysis with Sufficient Conditions

Here, a Separation Theorem about K-Independent Subspace Analysis (K real or complex), a generalization of K-Independent Component Analysis (KICA) is proven. According to the theorem, KISA estimation can be executed in two steps under certain conditions. In the first step, 1-dimensional KICA estimation is executed. In the second step, optimal permutation of the KICA elements is searched for. We present sufficient conditions for the KISA Separation Theorem. Namely, we shall show that (i) spherically symmetric sources (both for real and complex cases), as well as (ii) real 2-dimensional sources invariant to 90 degree rotation, among others, satisfy the conditions of the theorem.

math.ST↗

arXiv2006

Separation Theorem for Independent Subspace Analysis with Sufficient Conditions

Here, a separation theorem about Independent Subspace Analysis (ISA), a generalization of Independent Component Analysis (ICA) is proven. According to the theorem, ISA estimation can be executed in two steps under certain conditions. In the first step, 1-dimensional ICA estimation is executed. In the second step, optimal permutation of the ICA elements is searched for. We present sufficient conditions for the ISA Separation Theorem. Namely, we shall show that (i) elliptically symmetric sources, (ii) 2-dimensional sources invariant to 90 degree rotation, among others, satisfy the conditions of the theorem.

math.ST↗

arXiv2004

Intelligent encoding and economical communication in the visual stream

The theory of computational complexity is used to underpin a recent model of neocortical sensory processing. We argue that encoding into reconstruction networks is appealing for communicating agents using Hebbian learning and working on hard combinatorial problems, which are easy to verify. Computational definition of the concept of intelligence is provided. Simulations illustrate the idea.

q-bio.NC↗

arXiv2003

Centralized reward system gives rise to fast and efficient work sharing for intelligent Internet agents lacking direct communication

WWW has a scale-free structure where novel information is often difficult to locate. Moreover, Intelligent agents easily get trapped in this structure. Here a novel method is put forth, which turns these traps into information repositories, supplies: We populated an Internet environment with intelligent news foragers. Foraging has its associated cost whereas foragers are rewarded if they detect not yet discovered novel information. The intelligent news foragers crawl by using the estimated long-term cumulated reward, and also have a finite sized memory: the list of most promising supplies. Foragers form an artificial life community: the most successful ones are allowed to multiply, while unsuccessful ones die out. The specific property of this community is that there is no direct communication amongst foragers but the centralized rewarding system. Still, fast division of work is achieved.

cs.IR↗

arXiv2003

Controlled hierarchical filtering: Model of neocortical sensory processing

A model of sensory information processing is presented. The model assumes that learning of internal (hidden) generative models, which can predict the future and evaluate the precision of that prediction, is of central importance for information extraction. Furthermore, the model makes a bridge to goal-oriented systems and builds upon the structural similarity between the architecture of a robust controller and that of the hippocampal entorhinal loop. This generative control architecture is mapped to the neocortex and to the hippocampal entorhinal loop. Implicit memory phenomena; priming and prototype learning are emerging features of the model. Mathematical theorems ensure stability and attractive learning properties of the architecture. Connections to reinforcement learning are also established: both the control network, and the network with a hidden model converge to (near) optimal policy under suitable conditions. Falsifying predictions, including the role of the feedback connections between neocortical areas are made.

cs.NE↗

arXiv2003

Emergence of scale-free properties in Hebbian networks

The fundamental `plasticity' of the nervous system (i.e high adaptability at different structural levels) is primarily based on Hebbian learning mechanisms that modify the synaptic connections. The modifications rely on neural activity and assign a special dynamic behavior to the neural networks. Another striking feature of the nervous system is that spike based information transmission, which is supposed to be robust against noise, is noisy in itself: the variance of the spiking of the individual neurons is surprisingly large which may deteriorate the adequate functioning of the Hebbian mechanisms. In this paper we focus on networks in which Hebbian-like adaptation is induced only by external random noise and study spike-timing dependent synaptic plasticity. We show that such `HebbNets' are able to develop a broad range of network structures, including scale-free small-world networks. The development of such network structures may provide an explanation of the role of noise and its interplay with Hebbian plasticity. We also argue that this model can be seen as a unification of the famous Watts-Strogatz and preferential attachment models of small-world nets.

nlin.AO↗

arXiv2003

Towards a theory of consciousness: Proposal for the resolution of the homunculus fallacy with predictions

In this paper we argue that no forms of Turing test are either necessary or sufficient to establish if a machine is conscious or not. Furthermore, from a modeling point of view, the problem is that the Turing test does not really provide testable predictions. We believe that the model structure should explain the function (of consciousness). We argue that the cornerstone of any model on consciousness is to (partly) overcome the obstacle of the homunculus fallacy about the use of representations. In this contribution a possible solution is suggested, which makes use of reflexive architectures. The emerging computational constraints on the proposed architecture have lead to testable predictions on the dynamical behavior of the biological substrate. Interestingly, these predictions are in good agreement with recent experimental observations.

nlin.AO↗

arXiv2003

Kalman-filtering using local interactions

There is a growing interest in using Kalman-filter models for brain modelling. In turn, it is of considerable importance to represent Kalman-filter in connectionist forms with local Hebbian learning rules. To our best knowledge, Kalman-filter has not been given such local representation. It seems that the main obstacle is the dynamic adaptation of the Kalman-gain. Here, a connectionist representation is presented, which is derived by means of the recursive prediction error method. We show that this method gives rise to attractive local learning rules and can adapt the Kalman-gain.

cs.AI↗

arXiv2003

Temporal plannability by variance of the episode length

Optimization of decision problems in stochastic environments is usually concerned with maximizing the probability of achieving the goal and minimizing the expected episode length. For interacting agents in time-critical applications, learning of the possibility of scheduling of subtasks (events) or the full task is an additional relevant issue. Besides, there exist highly stochastic problems where the actual trajectories show great variety from episode to episode, but completing the task takes almost the same amount of time. The identification of sub-problems of this nature may promote e.g., planning, scheduling and segmenting Markov decision processes. In this work, formulae for the average duration as well as the standard deviation of the duration of events are derived. The emerging Bellman-type equation is a simple extension of Sobel's work (1982). Methods of dynamic programming as well as methods of reinforcement learning can be applied for our extension. Computer demonstration on a toy problem serve to highlight the principle.

cs.AI↗

arXiv2003

Kalman filter control in the reinforcement learning framework

There is a growing interest in using Kalman-filter models in brain modelling. In turn, it is of considerable importance to make Kalman-filters amenable for reinforcement learning. In the usual formulation of optimal control it is computed off-line by solving a backward recursion. In this technical note we show that slight modification of the linear-quadratic-Gaussian Kalman-filter model allows the on-line estimation of optimal control and makes the bridge to reinforcement learning. Moreover, the learning rule for value estimation assumes a Hebbian form weighted by the error of the value estimation.

cs.LG↗

arXiv2002

Searching for Plannable Domains can Speed up Reinforcement Learning

Reinforcement learning (RL) involves sequential decision making in uncertain environments. The aim of the decision-making agent is to maximize the benefit of acting in its environment over an extended period of time. Finding an optimal policy in RL may be very slow. To speed up learning, one often used solution is the integration of planning, for example, Sutton's Dyna algorithm, or various other methods using macro-actions. Here we suggest to separate plannable, i.e., close to deterministic parts of the world, and focus planning efforts in this domain. A novel reinforcement learning method called plannable RL (pRL) is proposed here. pRL builds a simple model, which is used to search for macro actions. The simplicity of the model makes planning computationally inexpensive. It is shown that pRL finds an optimal policy, and that plannable macro actions found by pRL are near-optimal. In turn, it is unnecessary to try large numbers of macro actions, which enables fast learning. The utility of pRL is demonstrated by computer simulations.

cs.AI↗