Utilizing Novelty-Based Evolution Strategies to Train Transformers in Reinforcement Learning

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.

**Read paper on the following link:** https://ifaamas.org/Proceedings/aamas2022/pdfs/p1237.pdf **Abstract:** Advances in Reinforcement Learning (RL) have demonstrated data efficiency and optimal control over large state spaces at the cost of scalable performance. Genetic methods, on the other hand, provide scalability but depict hyperparameter sensitivity towards evolutionary operations. A combination of the two methods has recently demonstrated success in scaling RL agents to high-dimensional action spaces. Parallel to recent developments, we present the Evolution-based Soft Actor-Critic (ESAC), a scalable RL algorithm. We abstract exploration from exploitation by combining Evolution Strategies (ES) with Soft Actor-Critic (SAC). Through this lens, we enable dominant skill transfer between offsprings by making use of soft winner selections and genetic crossovers in hindsight. Simultaneously we improve hyperparameter sensitivity in evolutions using the novel Automatic Mutation Tuning (AMT). AMT gradually replaces the entropy framework of SAC allowing the population to succeed at the task while acting as randomly as possible, without making use of backpropagation updates. In a study of challenging locomotion tasks consisting of high-dimensional action spaces and sparse rewards, ESAC demonstrates improved performance and sample efficiency in comparison to the Maximum Entropy framework. Additionally, ESAC presents efficacious use of hardware resources and algorithm overhead.

Game agents are efficiently implemented through different AI techniques, such as neural network, reinforcement learning, and evolutionary search. Although there are many works for each approach, we present a critical analysis and comparison between them, suggesting a common benchmark and parameter configurations. The evolutionary strategy implements the NeuroEvolution of Augmenting Topologies algorithm, while the reinforcement learning agent leverages Q-Learning and Proximal Policy Optimization. We formulate and empirically compare this set of solutions using the Flappy Bird game as a test scenario. We also compare different representations of state and reward functions for each method. All methods were able to generate agents that can play the game, where the NEAT algorithm had the best results, reaching the goal of never losing.

Evolution Strategies (ESs) have recently become popular for training deep neural networks, in particular on reinforcement learning tasks, a special form of controller design. Compared to classic problems in continuous direct search, deep networks pose extremely high-dimensional optimization problems, with many thousands or even millions of variables. In addition, many control problems give rise to a stochastic fitness function. Considering the relevance of the application, we study the suitability of evolution strategies for high-dimensional, stochastic problems. Our results give insights into which algorithmic mechanisms of modern ES are of value for the class of problems at hand, and they reveal principled limitations of the approach. They are in line with our theoretical understanding of ESs. We show that combining ESs that offer reduced internal algorithm cost with uncertainty handling techniques yields promising methods for this class of problems.

Evolution Strategies (ES) have emerged as a competitive alternative for model-free reinforcement learning, showcasing exemplary performance in tasks like Mujoco and Atari. Notably, they shine in scenarios with imperfect reward functions, making them invaluable for real-world applications where dense reward signals may be elusive. Yet, an inherent assumption in ES—that all input features are task-relevant—poses challenges, especially when confronted with irrelevant features common in real-world problems. This work scrutinizes this limitation, particularly focusing on the Natural Evolution Strategies (NES) variant. We propose NESHT, a novel approach that integrates Hard-Thresholding (HT) with NES to champion sparsity, ensuring only pertinent features are employed. Backed by rigorous analysis and empirical tests, NESHT demonstrates its promise in mitigating the pitfalls of irrelevant features and shines in complex decision-making problems like noisy Mujoco and Atari tasks. Our code is available at https://github.com/cangcn/NES-HT.

Randomized direct search algorithms for continuous domains, such as evolution strategies, are basic tools in machine learning. They are especially needed when the gradient of an objective function (e.g., loss, energy, or reward function) cannot be computed or estimated efficiently. Application areas include supervised and reinforcement learning as well as model selection. These randomized search strategies often rely on normally distributed additive variations of candidate solutions. In order to efficiently search in non-separable and ill-conditioned landscapes the covariance matrix of the normal distribution must be adapted, amounting to a variable metric method. Consequently, covariance matrix adaptation (CMA) is considered state-of-the-art in evolution strategies. In order to sample the normal distribution, the adapted covariance matrix needs to be decomposed, requiring in general Θ(n 3) operations, where n is the search space dimension. We propose a new update mechanism which can replace a rank-one covariance matrix update and the computationally expensive decomposition of the covariance matrix. The newly developed update rule reduces the computational complexity of the rank-one covariance matrix adaptation to Θ(n 2) without resorting to outdated distributions. We derive new versions of the elitist covariance matrix adaptation evolution strategy (CMA-ES) and the multi-objective CMA-ES. These algorithms are equivalent to the original procedures except that the update step for the variable metric distribution scales better in the problem dimension. We also introduce a simplified variant of the non-elitist CMA-ES with the incremental covariance matrix update and investigate its performance. Apart from the reduced time-complexity of the distribution update, the algebraic computations involved in all new algorithms are simpler compared to the original versions. The new update rule improves the performance of the CMA-ES for large scale machine learning problems in which the objective function can be evaluated fast.

Novel method of mobile edge computation offloading based on evolutionary game strategy for IoT devices

Verification and validation (V&V) is a critical issue in both multi-agent systems (MAS) and agent-based social simulation (ABSS). As the first step towards V&V methods for MAS and ABSS, this paper investigates whether different computational models can produce the same results. Specifically, we compare three computational models with different learning mechanisms in a multiagent-based simulation and analyze the results of these models in a bargaining game as one of the fundamental examples in game theory. This type of V&V is not based on the between-models addressed in conventional research, but on a within-model. A comparison of the simulation results reveals that (1) computational models and simulation results are minimally verified and validated in the case of ES(evolutionary strategy)- and RL(reinforcement learning)-based agents; and (2) learning mechanisms that enable agents to acquire their rational behaviors differ according to the knowledge representation (i.e., the strategies in the bargaining game) of the agents.KeywordsVerification and validationmultiagent-based simulationcomparison of different modelslearning mechanismbargaining game

With the continuous development of the times, there have been many significant breakthroughs in the field of reinforcement learning, not only reflected in the improvement of operational efficiency, but also in the emergence of new methods, all of which contribute to the continuous concretization of reinforcement learning in reality. This article mainly introduces the current research status of reinforcement learning, the introduction of common basic reinforcement learning algorithms such as value function estimation, policy gradient method, and Actor Critic algorithm. In recent years, a review of transfer learning, width planning and active learning based methods Olive, Reptile algorithm, evolutionary strategy (ES), two self play training schemes (Chainer and Pool), and the application analysis of reinforcement learning in the fields of gaming, robot control, transportation, healthcare, finance, and energy is conducted. This article aims to provide a review of the current research status in recent years, in order to provide a reference for the future development of reinforcement learning and indicate the existing problems in current research.

Maintaining the stability of the modern power grid is becoming increasingly difficult due to fluctuating power consumption, unstable power supply coming from renewable energies, and unpredictable accidents such as man-made and natural disasters. As the operation on the power grid must consider its impact on future stability, reinforcement learning (RL) has been employed to provide sequential decision-making in power grid management. However, existing methods have not considered the environmental constraints. As a result, the learned policy has risk of selecting actions that violate the constraints in emergencies, which will escalate the issue of overloaded power lines and lead to large-scale blackouts. In this work, we propose a novel method for this problem, which builds on top of the search-based planning algorithm. At the planning stage, the search space is limited to the action set produced by the policy. The selected action strictly follows the constraints by testing its outcome with the simulation function provided by the system. At the learning stage, to address the problem that gradients cannot be propagated to the policy, we introduce Evolutionary Strategies (ES) with black-box policy optimization to improve the policy directly, maximizing the returns of the long run. In NeurIPS 2020 Learning to Run Power Network (L2RPN) competition, our solution safely managed the power grid and ranked first in both tracks.

Many real-world problems involve cooperation and/or competition among multiple agents. These problems often can be formulated as multi-agent problems. Recently, Reinforcement Learning (RL) has made significant progress on single-agent problems. However, multi-agent problems still cannot be easily solved by traditional RL algorithms. First, the multi-agent environment is considered as a non-stationary system. Second, most multi-agent environments only provide a shared team reward as feedback. As a result, agents may not be able to learn proper cooperative or competitive behaviors by traditional RL. Our algorithm adopts Evolution Strategies (ES) for optimizing policy which is used to control agents and a value decomposition method for estimating proper fitness for each policy. Evolutionary Algorithm is considered as a promising alternative for signal-agent problems. Owing to its simplicity, scalability, and efficiency on zeroth-order optimization, EAs can even outperform RLs on some tasks. In order to solve multi-agent problems by EA, a value decomposition method is used to decompose the team reward. Our method is parallel on multiple cores, which can speed up our algorithm significantly. We test our algorithm on two benchmarking environments, and the experiment results show that our algorithm is better than traditional RL and other representative gradient-free methods.

Uncertainty arises in reinforcement learning from various sources, and therefore it is necessary to consider statistics based on several roll-outs for evaluating behavioral policies. We add an adaptive uncertainty handling based on Hoeffding and empirical Bernstein races to the CMA-ES, a variable metric evolution strategy proposed for direct policy search. The uncertainty handling adjusts individually the number of episodes considered for the evaluation of a policy. The performance estimation is kept just accurate enough for a sufficiently good ranking of candidate policies, which is in turn sufficient for the CMA-ES to find better solutions. This increases the learning speed as well as the robustness of the algorithm.

Intrinsically, driving is a Markov Decision Process which suits well the reinforcement learning paradigm. In this paper, we propose a novel agent which learns to drive a vehicle without any human assistance. We use the concept of reinforcement learning and evolutionary strategies to train our agent in a 2D simulation environment. Our model's architecture goes beyond the World Model's by introducing difference images in the auto encoder. This novel involvement of difference images in the auto-encoder gives better representation of the latent space with respect to the motion of vehicle and helps an autonomous agent to learn more efficiently how to drive a vehicle. Results show that our method requires fewer (96% less) total agents, (87.5% less) agents per generations, (70% less) generations and (90% less) rollouts than the original architecture while achieving the same accuracy of the original.

Self-Preserving Genetic Algorithms (SPGA) combine the evolutionary strategy of a genetic algorithm with safety assurance methods commonly implemented in safe reinforcement learning (SRL), a branch of reinforcement learning (RL) that accounts for safety in the exploration and decision-making process of the agent. Safe learning approaches are especially important in safety-critical environments, where failure to account for the safety of the controlled system could result in the loss of millions of dollars in hardware or bodily harm to people working nearby, as is true of many cyber-physical systems. While SRL is a viable approach to safe learning, there are many challenges that must be taken into consideration when training agents, such as sample efficiency, stability, and exploration---an issue that is easily addressed by the evolutionary strategy of a genetic algorithm. By combining GAs with the safety mechanisms used with SRL, SPGA offers a safe learning alternative that is able to explore large areas of the solution space, addressing SRL's challenge of exploration. This work implements SPGA with both action masking and run time assurance safety strategies to evolve safe controllers for three types of discrete action space environments applicable to cyber physical systems (control, routing, and operations) and under various safety conditions. Training and testing evaluation metrics are compared with results from SRL trained controllers to validate results. SPGA and SRL controllers are trained across 5 random seeds and evaluated on 500 episodes to calculate average wall time to train, average expected return, and percentage of safe action evaluation metrics. SPGA achieves comparable reward and safety performance results with significantly improved training efficiency (55x faster on average), demonstrating the effectiveness of this safe learning approach.