Challenges in High-Dimensional Reinforcement Learning with Evolution Strategies

In computer science, reinforcement learning is a powerful framework with which artificial agents can learn to maximize their performance for any given Markov decision process (MDP). Advances over the last decade, in combination with deep neural networks, have enjoyed performance advantages over humans in many difficult task settings. However, such frameworks perform far less favorably when evaluated in their ability to generalize or transfer representations across different tasks. Existing algorithms that facilitate transfer typically are limited to cases in which the transition function or the optimal policy is portable to new contexts, but achieving “deep transfer” characteristic of human behavior has been elusive. Such transfer typically requires discovery of abstractions that permit analogical reuse of previously learned representations to superficially distinct tasks. Here, we demonstrate that abstractions that minimize error in predictions of reward outcomes generalize across tasks with different transition and reward functions. Such reward-predictive representations compress the state space of a task into a lower dimensional representation by combining states that are equivalent in terms of both the transition and reward functions. Because only state equivalences are considered, the resulting state representation is not tied to the transition and reward functions themselves and thus generalizes across tasks with different reward and transition functions. These results contrast with those using abstractions that myopically maximize reward in any given MDP and motivate further experiments in humans and animals to investigate if neural and cognitive systems involved in state representation perform abstractions that facilitate such equivalence relations.

Evolution strategies have been demonstrated to have the strong ability to roughly train deep neural networks and well accomplish reinforcement learning tasks. However, existing evolution strategies designed specially for deep reinforcement learning only involve the plain variants which can not realize the adaptation of mutation strength or other advanced techniques. The research of applying advanced and effective evolution strategies to reinforcement learning in an efficient way is still a gap. To this end, this paper proposes a restart-based rank-1 evolution strategy for reinforcement learning. When training the neural network, it adapts the mutation strength and updates the principal search direction in a way similar to the momentum method, which is an ameliorated version of stochastic gradient ascent. Besides, two mechanisms, i.e., the adaptation of the number of elitists and the restart procedure, are integrated to deal with the issue of local optima. Experimental results on classic control problems and Atari games show that the proposed algorithm is superior to or competitive with state-of-the-art algorithms for reinforcement learning, demonstrating the effectiveness of the proposed algorithm.

Evolution strategies (ESs), as a family of black-box optimization algorithms, recently emerge as a scalable alternative to reinforcement learning (RL) approaches such as Q-learning or policy gradient and are much faster when many central processing units (CPUs) are available due to better parallelization. In this article, we propose a systematic incremental learning method for ES in dynamic environments. The goal is to adjust previously learned policy to a new one incrementally whenever the environment changes. We incorporate an instance weighting mechanism with ES to facilitate its learning adaptation while retaining scalability of ES. During parameter updating, higher weights are assigned to instances that contain more new knowledge, thus encouraging the search distribution to move toward new promising areas of parameter space. We propose two easy-to-implement metrics to calculate the weights: instance novelty and instance quality. Instance novelty measures an instance's difference from the previous optimum in the original environment, while instance quality corresponds to how well an instance performs in the new environment. The resulting algorithm, instance weighted incremental evolution strategies (IW-IESs), is verified to achieve significantly improved performance on challenging RL tasks ranging from robot navigation to locomotion. This article thus introduces a family of scalable ES algorithms for RL domains that enables rapid learning adaptation to dynamic environments.

Self-play has been shown to be effective to provide a proper training curriculum for a reinforcement learning agent in competitive multi-agent environments without direct supervision. However, its performance is still unstable for problems with sparse rewards, e.g., the scoring task with goalkeeper for robots in RoboCup soccer. It is challenging to solve these tasks in reinforcement learning, especially for those that require combining high-level actions with flexible control. To address these challenges, we introduce a distributed self-play training framework for an extended proximal policy optimization (PPO) algorithm that learns to act in parameterized action space and plays against a group of opponents, i.e., a league. Experiments on the domain of simulated RoboCup soccer show that, the approach is effective and learns more robust policies against various opponents compared to existing reinforcement learning methods. A demonstration video is available online at https://youtu.be/BuLli1vND4.

We propose reinforcement learning on simple networks consisting of random connections of spiking neurons (both recurrent and feed-forward) that can learn complex tasks with very little trainable parameters. Such sparse and randomly interconnected recurrent spiking networks exhibit highly non-linear dynamics that transform the inputs into rich high-dimensional representations based on the current and past context. The random input representations can be efficiently interpreted by an output (or readout) layer with trainable parameters. Systematic initialization of the random connections and training of the readout layer using Q-learning algorithm enable such small random spiking networks to learn optimally and achieve the same learning efficiency as humans on complex reinforcement learning (RL) tasks like Atari games. In fact, the sparse recurrent connections cause these networks to retain fading memory of past inputs, thereby enabling them to perform temporal integration across successive RL time-steps and learn with partial state inputs. The spike-based approach using small random recurrent networks provides a computationally efficient alternative to state-of-the-art deep reinforcement learning networks with several layers of trainable parameters.

Background: Traditional supervised learning (SL) assumes data points are independently and identically distributed (i.i.d.), which overlooks dependencies in real-world data. Reinforcement learning (RL), in contrast, models dependencies through state transitions. Objectives: This study aims to bridge SL and RL by reformulating SL problems as RL tasks, enabling the application of RL techniques to a wider range of SL scenarios. We aim to model SL data as interconnected, and develop novel temporal difference (TD) algorithms that can accommodate diverse data types. Our objectives are to (1) establish conditions where TD outperforms ordinary least squares (OLS), (2) provide convergence guarantees for the generalized TD algorithm, and (3) validate the approach empirically using synthetic and real-world datasets. Methods: We reformulate traditional SL as a RL problem by modeling data points as a Markov Reward Process (MRP). We then introduce a concept analogous to the inverse link function in generalized linear models, allowing our TD algorithm to handle various data types. Our analysis, grounded in variance estimation, identifies conditions where TD outperforms OLS. We establish a convergence guarantee by conceptualizing the TD update rule as a generalized Bellman operator. Empirical validation begins with synthetic data progressively matching theoretical assumptions to verify our analysis, followed by evaluations on real-world datasets to demonstrate practical utility. Results: Our theoretical analysis shows that TD can outperform OLS in estimation accuracy when data noise is correlated. Our approach generalizes across various loss functions and SL datasets. We prove that the Bellman operator in our TD framework is a contraction, ensuring convergence for both expected and stochastic TD updates. Empirically, TD outperforms SL baselines when data aligns with its assumptions, remains competitive across diverse datasets, and is robust to hyperparameter choices. Conclusions: This study demonstrates that SL can be reformulated as a problem of interconnected data modeled by an MRP, effectively solved using TD learning. Our generalized TD is theoretically sound, with convergence guarantees, and practically effective. It generalizes OLS, offering superior performance on correlated data. This work enables RL techniques to benefit SL tasks, offering a pathway for future advancements.

In recent years, deep neural networks (DNNs) have achieved great successes in many areas, such as cognitive computation, pattern recognition, and computer vision. Although many hand-crafted deep networks have been proposed in the literature, designing a well-behaved neural network for a specific application requires high-level expertise yet. Hence, the automatic architecture design of DNNs has become a challenging and important problem. In this paper, we propose a new reinforcement learning method, whose action policy is to select neural blocks and construct deep networks. We define the action search space with three types of neural blocks, i.e., dense block, residual block, and inception-like block. Additionally, we have also designed several variants for the residual and inception-like blocks. The optimal network is automatically learned by a Q-learning agent, which is iteratively trained to generate well-performed deep networks. To evaluate the proposed method, we have conducted experiments on three datasets, MNIST, SVHN, and CIFAR-10, for image classification applications. Compared with existing hand-crafted and auto-generated neural networks, our auto-designed neural network delivers promising results. Moreover, the proposed reinforcement learning algorithm for deep networks design only runs on one GPU, demonstrating much higher efficiency than most of the previous deep network search approaches.

Reinforcement learning (RL) tasks are challenging to implement, execute and test due to algorithmic instability, hyper-parameter sensitivity, and heterogeneous distributed communication patterns. We argue for the separation of logical component composition, backend graph definition, and distributed execution. To this end, we introduce RLgraph, a library for designing and executing reinforcement learning tasks in both static graph and define-by-run paradigms. The resulting implementations are robust, incrementally testable, and yield high performance across different deep learning frameworks and distributed backends.

Maternal immune activation (MIA) during pregnancy increases the risk for the unborn foetus to develop neurodevelopmental conditions such as autism spectrum disorder and schizophrenia later in life. MIA mouse models recapitulate behavioural and biological phenotypes relevant to both conditions, and are valuable models to test novel treatment approaches. Selenium (Se) has potent anti-inflammatory properties suggesting it may be an effective prophylactic treatment against MIA. The aim of this study was to determine if Se supplementation during pregnancy can prevent adverse effects of MIA on offspring brain and behaviour in a mouse model.Selenium was administered via drinking water (1.5 ppm) to pregnant dams from gestational day (GD) 9 to birth, and MIA was induced at GD17 using polyinosinic:polycytidylic acid (poly-I:C, 20 mg/kg via intraperitoneal injection). Foetal placenta and brain cytokine levels were assessed using a Luminex assay and brain elemental nutrients assessed using inductively coupled plasma- mass spectrometry. Adult offspring were behaviourally assessed using a reinforcement learning paradigm, the three-chamber sociability test and the open field test. MIA elevated placental IL-1β and IL-17, and Se supplementation successfully prevented this elevation. MIA caused an increase in foetal brain calcium, which was prevented by Se supplement. MIA caused in offspring a female-specific reduction in sociability, which was recovered by Se, and a male-specific reduction in social memory, which was not recovered by Se. Exposure to poly-I:C or selenium, but not both, reduced performance in the reinforcement learning task. Computational modelling indicated that this was predominantly due to increased exploratory behaviour, rather than reduced rate of learning the location of the food reward.This study demonstrates that while Se may be beneficial in ameliorating sociability deficits caused by MIA, it may have negative effects in other behavioural domains. Caution in the use of Se supplementation during pregnancy is therefore warranted.

We previously proposed a method (1998) for acquiring a good evaluation function of the game Othello based on min-max reinforcement learning. In the previous work, we employed a gradient descent method for training the normalized Gaussian network (NGnet) that represents the Othello's evaluation function. However, a lot of games were required to train the NGnet. In this study, we employ the previously proposed online EM algorithm to train the NGnet. The online EM algorithm converges faster than the gradient descent method, and it is suitable for dynamic environments, such as in reinforcement learning tasks. In this article, we introduce a new architecture that is composed of a certain number of NGnets to represent the evaluation function. In this architecture, each of the NGnets is independently trained by the online EM algorithm. Our experiments show that a good evaluation function can be obtained by this new architecture through a smaller number of training games than by the previous scheme.

This paper revisits the temporal difference (TD) learning algorithm for the policy evaluation tasks in reinforcement learning. Typically, the performance of TD(0) and TD( λ) is very sensitive to the choice of stepsizes. Oftentimes, TD(0) suffers from slow convergence. Motivated by the tight link between the TD(0) learning algorithm and the stochastic gradient methods, we develop a provably convergent adaptive projected variant of the TD(0) learning algorithm with linear function approximation that we term AdaTD(0). In contrast to the TD(0), AdaTD(0) is robust or less sensitive to the choice of stepsizes. Analytically, we establish that to reach an ϵ accuracy, the number of iterations needed is [Formula: see text] in the general case, where ρ represents the speed of the underlying Markov chain converges to the stationary distribution. This implies that the iteration complexity of AdaTD(0) is no worse than that of TD(0) in the worst case. When the stochastic semi-gradients are sparse, we provide theoretical acceleration of AdaTD(0). Going beyond TD(0), we develop an adaptive variant of TD( λ), which is referred to as AdaTD( λ). Empirically, we evaluate the performance of AdaTD(0) and AdaTD( λ) on several standard reinforcement learning tasks, which demonstrate the effectiveness of our new approaches.

Both emotional facial expressions and markers of racial-group belonging are ubiquitous signals in social interaction, but little is known about how these signals together affect future behavior through learning. To address this issue, we investigated how emotional (threatening or friendly) in-group and out-group faces reinforced behavior in a reinforcement-learning task. We asked whether reinforcement learning would be modulated by intergroup attitudes (i.e., racial bias). The results showed that individual differences in racial bias critically modulated reinforcement learning. As predicted, racial bias was associated with more efficiently learned avoidance of threatening out-group individuals. We used computational modeling analysis to quantitatively delimit the underlying processes affected by social reinforcement. These analyses showed that racial bias modulates the rate at which exposure to threatening out-group individuals is transformed into future avoidance behavior. In concert, these results shed new light on the learning processes underlying social interaction with racial-in-group and out-group individuals.

Computational models of the basal ganglia (BG) provide a mechanistic account of different phenomena observed during reinforcement learning tasks performed by healthy individuals, as well as by patients with various nervous or mental disorders. The aim of the present work was to develop a BG model that could represent a good compromise between simplicity and completeness. Based on more complex (fine-grained neural network, FGNN) models, we developed a new (coarse-grained neural network, CGNN) model by replacing layers of neurons with single nodes that represent the collective behavior of a given layer while preserving the fundamental anatomical structures of BG. We then compared the functionality of both the FGNN and CGNN models with respect to several reinforcement learning tasks that are based on BG circuitry, such as the Probabilistic Selection Task, Probabilistic Reversal Learning Task and Instructed Probabilistic Selection Task. We showed that CGNN still has a functionality that mirrors the behavior of the most often used reinforcement learning tasks in human studies. The simplification of the CGNN model reduces its flexibility but improves the readability of the signal flow in comparison to more detailed FGNN models and, thus, can help to a greater extent in the translation between clinical neuroscience and computational modeling.

A reinforcement learning algorithm is proposed that can cope with high dimensionality for a class of problems with symmetrical actions. The action selection does not need considering all the states but only needs looking at a part of the states. Moreover, every symmetrical action is related to the same kind of part of state, and thus the value function can be shared, which greatly reduces the reinforcement learning problem size. The overall learning algorithm is equivalent to the standard reinforcement learning algorithm. Simulation results and other aspects are compared with standard and other reinforcement learning algorithms. Reduction in dimensionality, much faster convergence without worsening other objectives show the effectiveness of the proposed mechanism on a high dimensional optimization problem having symmetrical actions.

RL-VAEGAN: Adversarial defense for reinforcement learning agents via style transfer