Search

Perceptual aliasing challenges reinforcement learning agents. They struggle to learn stable policies by failing to identify and disambiguate perceptually identical states in the environment that require different actions to reach a goal. As the agent often has only a local frame of reference, it cannot represent the global environment. Frame-of-reference-based learning is a feature of vertebrate intelligence that allows multiple simultaneous representations of an environment at different levels of abstraction. This enables the resolution of patterns that are made up of patterns that are made up of features. The evolutionary computation technique of learning classifier systems has shown promise in learning nested patterns in single-step domains. This work uses the frame-of-reference concept within a learning classifier system to learn stable policies in non-Markov multistep domains. Considering aliased states at a constituent level enables the system to place them appropriately in holistic-level policies. Instead of enumerating a huge search space, evolution computation empowers the novel system to evolve fitter rules and policies. The experimental results show that the novel system effectively solves complex aliasing patterns in non-Markov environments that have been challenging to artificial agents. For example, the novel system utilizes only 6.5, 3.71, and 3.22 steps to resolve Maze10, Littman57, and Woods102, respectively.

Modern classifier systems can effectively classify targets that consist of simple patterns. However, they can fail to detect hierarchical patterns of features that exist in many real-world problems, such as understanding speech or recognizing object ontologies. Biological nervous systems have the ability to abstract knowledge from simple and small-scale problems in order to then apply it to resolve more complex problems in similar and related domains. It is thought that lateral asymmetry of biological brains allows modular learning to occur at different levels of abstraction, which can then be transferred between tasks. This work develops a novel evolutionary machine-learning (EML) system that incorporates lateralization and modular learning at different levels of abstraction. The results of analyzable Boolean tasks show that the lateralized system has the ability to encapsulate underlying knowledge patterns in the form of building blocks of knowledge (BBK). Lateralized abstraction transforms complex problems into simple ones by reusing general patterns (e.g., any parity problem becomes a sequence of the 2-bit parity problem). By enabling abstraction in evolutionary computation, the lateralized system is able to identify complex patterns (e.g., in hierarchical multiplexer (HMux) problems) better than existing systems.

Biological nervous systems can learn knowledge from simple and small-scale problems and then apply it to resolve more complex and large-scale problems in similar and related domains. However, the rudimentary attempts to apply this transfer learning in artificial intelligence systems have struggled. This is maybe due to the homogeneous nature of their knowledge representation. It is believed that it is the lateral asymmetry of the brain, enabling modular learning at different levels of abstraction, which facilitates transfer between tasks. Learning classifier systems (LCSs) are a rule-based evolutionary computation technique that automatically clusters inputs into environmental niches, which makes it an ideal candidate for implementing lateralization. Recently LCSs based systems have applied lateralization and modular learning at different levels of abstraction to solve complex problems in Boolean, computer vision, and navigation domains. This paper aims to bring these three separate implementations together for the first time to understand the methodology of lateralization and appreciate its benefits. The experimental results demonstrate that the LCSs based lateralized systems outperformed state-of-the-art homogeneous systems in solving complex problems. The advances arise from the ability to consider input at both the constituent level and holistic level simultaneously, such that the most appropriate viewpoint controls the system.

Deep learning is an important field of machine learning. It is playing a critical role in a variety of applications ranging from self-driving cars to security and surveillance. However, deep networks have deep flaws. For example, they are highly vulnerable to adversarial attacks. One reason may be the homogeneous nature of their knowledge representation, which allows a single disruptive pattern to cause miss-classification. Biological intelligence has lateral asymmetry, which allows heterogeneous, modular learning at different levels of abstraction, enabling different representations of the same object. This work aims to incorporate lateralization and modular learning at different levels of abstraction in an evolutionary machine learning system. The results of image classification tasks show that the lateralized system efficiently learns hierarchical distributions of knowledge, demonstrating performance that is similar to (or better than) other state-of-the-art deep systems as it reasons using multiple representations. Crucially, the novel system outperformed all the state-of-the-art deep models for the classification of normal and adversarial images by 0.43% - 2.56% and 2.15% - 25.84%, respectively. Lateralisation enabled the system to exhibit robustness beyond previous work, which advocates for the creation of data sets that enable components of objects and the objects themselves to be learned specifically or in an end-to-end manner.