Improved robustness of reinforcement learning policies upon conversion to spiking neuronal network platforms applied to Atari Breakout game

Towards biologically plausible model-based reinforcement learning in recurrent spiking networks by dreaming new experiences

Humans and animals can learn new skills after practicing for a few hours, while current reinforcement learning algorithms require a large amount of data to achieve good performances. Recent model-based approaches show promising results by reducing the number of necessary interactions with the environment to learn a desirable policy. However, these methods require biological implausible ingredients, such as the detailed storage of older experiences, and long periods of offline learning. The optimal way to learn and exploit world-models is still an open question. Taking inspiration from biology, we suggest that dreaming might be an efficient expedient to use an inner model. We propose a two-module (agent and model) spiking neural network in which "dreaming" (living new experiences in a model-based simulated environment) significantly boosts learning. Importantly, our model does not require the detailed storage of experiences, and learns online the world-model and the policy. Moreover, we stress that our network is composed of spiking neurons, further increasing the biological plausibility and implementability in neuromorphic hardware.

Human-Level Control Through Directly Trained Deep Spiking Q-Networks

As the third-generation neural networks, spiking neural networks (SNNs) have great potential on neuromorphic hardware because of their high energy efficiency. However, deep spiking reinforcement learning (DSRL), that is, the reinforcement learning (RL) based on SNNs, is still in its preliminary stage due to the binary output and the nondifferentiable property of the spiking function. To address these issues, we propose a deep spiking Q -network (DSQN) in this article. Specifically, we propose a directly trained DSRL architecture based on the leaky integrate-and-fire (LIF) neurons and deep Q -network (DQN). Then, we adapt a direct spiking learning algorithm for the DSQN. We further demonstrate the advantages of using LIF neurons in DSQN theoretically. Comprehensive experiments have been conducted on 17 top-performing Atari games to compare our method with the state-of-the-art conversion method. The experimental results demonstrate the superiority of our method in terms of performance, stability, generalization and energy efficiency. To the best of our knowledge, our work is the first one to achieve state-of-the-art performance on multiple Atari games with the directly trained SNN.

QC_SANE: Robust Control in DRL Using Quantile Critic With Spiking Actor and Normalized Ensemble

Recently introduced deep reinforcement learning (DRL) techniques in discrete-time have resulted in significant advances in online games, robotics, and so on. Inspired from recent developments, we have proposed an approach referred to as Quantile Critic with Spiking Actor and Normalized Ensemble (QC_SANE) for continuous control problems, which uses quantile loss to train critic and a spiking neural network (NN) to train an ensemble of actors. The NN does an internal normalization using a scaled exponential linear unit (SELU) activation function and ensures robustness. The empirical study on multijoint dynamics with contact (MuJoCo)-based environments shows improved training and test results than the state-of-the-art approach: population coded spiking actor network (PopSAN).

A brain-inspired theory of mind spiking neural network improves multi-agent cooperation and competition

During dynamic social interaction, inferring and predicting others' behaviors through theory of mind (ToM) is crucial for obtaining benefits in cooperative and competitive tasks. Current multi-agent reinforcement learning (MARL) methods primarily rely on agent observations to select behaviors, but they lack inspiration from ToM, which limits performance. In this article, we propose a multi-agent ToM decision-making (MAToM-DM) model, which consists of a MAToM spiking neural network (MAToM-SNN) module and a decision-making module. We design two brain-inspired ToM modules (Self-MAToM and Other-MAToM) to predict others' behaviors based on self-experience and observations of others, respectively. Each agent can adjust its behavior according to the predicted actions of others. The effectiveness of the proposed model has been demonstrated through experiments conducted in cooperative and competitive tasks. The results indicate that integrating the ToM mechanism can enhance cooperation and competition efficiency and lead to higher rewards compared with traditional MARL models.

Mixture of personality improved spiking actor network for efficient multi-agent cooperation

Adaptive multi-agent cooperation with especially unseen partners is becoming more challenging in multi-agent reinforcement learning (MARL) research, whereby conventional deep-learning-based algorithms suffer from the poor new-player-generalization problem, possibly caused by not considering theory-of-mind theory (ToM). Inspired by the ToM personality in cognitive psychology, where a human can easily resolve this problem by predicting others' intuitive personality first before complex actions, we propose a biologically-plausible algorithm named the mixture of personality (MoP) improved spiking actor network (SAN). The MoP module contains a determinantal point process to simulate the formation and integration of different personality types, and the SAN module contains spiking neurons for efficient reinforcement learning. The experimental results on the benchmark cooperative overcooked task showed that the proposed MoP-SAN algorithm could achieve higher performance for the paradigms with (learning) and without (generalization) unseen partners. Furthermore, ablation experiments highlighted the contribution of MoP in SAN learning, and some visualization analysis explained why the proposed algorithm is superior to some counterpart deep actor networks.

Cooperative modular reinforcement learning for large discrete action space problem

Deep reinforcement learning (DRL) has achieved remarkable results on high-dimension state tasks. However, it suffers in hard convergence and low sample efficiency when solving large discrete action space problems. To meet these challenges, we develop a cooperative modular reinforcement learning (CMRL) method to distributedly solve the problems with a large discrete action space. A general yet effective task decomposition method is proposed to decompose the complex decision task in a large action space into multiple decision sub-tasks in small action subsets, using a rule-based action division method. The CMRL method consisting of multiple Critic networks is proposed to settle the multiple sub-tasks, where each Critic network learns a decomposed value function to obtain the local optimal action in a sub-task. The global optimal action is cooperatively chosen by all local optimal actions. Moreover, we propose a new parallel training mechanism, which trains multiple Critic networks with different models and multi-data in parallel. Mathematical properties are proposed to analyze the rationality and superiority of CMRL. Four different simulation experiments are conducted to verify the generality and effectiveness of CMRL for large action space problems. The results show that CMRL has superior performance on training efficiency compared with classical and latest DRL methods while maintaining the accuracy of the solution.

Toward robust and scalable deep spiking reinforcement learning

Deep reinforcement learning (DRL) combines reinforcement learning algorithms with deep neural networks (DNNs). Spiking neural networks (SNNs) have been shown to be a biologically plausible and energy efficient alternative to DNNs. Since the introduction of surrogate gradient approaches that allowed to overcome the discontinuity in the spike function, SNNs can now be trained with the backpropagation through time (BPTT) algorithm. While largely explored on supervised learning problems, little work has been done on investigating the use of SNNs as function approximators in DRL. Here we show how SNNs can be applied to different DRL algorithms like Deep Q-Network (DQN) and Twin-Delayed Deep Deteministic Policy Gradient (TD3) for discrete and continuous action space environments, respectively. We found that SNNs are sensitive to the additional hyperparameters introduced by spiking neuron models like current and voltage decay factors, firing thresholds, and that extensive hyperparameter tuning is inevitable. However, we show that increasing the simulation time of SNNs, as well as applying a two-neuron encoding to the input observations helps reduce the sensitivity to the membrane parameters. Furthermore, we show that randomizing the membrane parameters, instead of selecting uniform values for all neurons, has stabilizing effects on the training. We conclude that SNNs can be utilized for learning complex continuous control problems with state-of-the-art DRL algorithms. While the training complexity increases, the resulting SNNs can be directly executed on neuromorphic processors and potentially benefit from their high energy efficiency.

Training spiking neuronal networks to perform motor control using reinforcement and evolutionary learning

Artificial neural networks (ANNs) have been successfully trained to perform a wide range of sensory-motor behaviors. In contrast, the performance of spiking neuronal network (SNN) models trained to perform similar behaviors remains relatively suboptimal. In this work, we aimed to push the field of SNNs forward by exploring the potential of different learning mechanisms to achieve optimal performance. We trained SNNs to solve the CartPole reinforcement learning (RL) control problem using two learning mechanisms operating at different timescales: (1) spike-timing-dependent reinforcement learning (STDP-RL) and (2) evolutionary strategy (EVOL). Though the role of STDP-RL in biological systems is well established, several other mechanisms, though not fully understood, work in concert during learning in vivo. Recreating accurate models that capture the interaction of STDP-RL with these diverse learning mechanisms is extremely difficult. EVOL is an alternative method and has been successfully used in many studies to fit model neural responsiveness to electrophysiological recordings and, in some cases, for classification problems. One advantage of EVOL is that it may not need to capture all interacting components of synaptic plasticity and thus provides a better alternative to STDP-RL. Here, we compared the performance of each algorithm after training, which revealed EVOL as a powerful method for training SNNs to perform sensory-motor behaviors. Our modeling opens up new capabilities for SNNs in RL and could serve as a testbed for neurobiologists aiming to understand multi-timescale learning mechanisms and dynamics in neuronal circuits.

Solving the spike feature information vanishing problem in spiking deep Q network with potential based normalization

Brain-inspired spiking neural networks (SNNs) are successfully applied to many pattern recognition domains. The SNNs-based deep structure has achieved considerable results in perceptual tasks, such as image classification and target detection. However, applying deep SNNs in reinforcement learning (RL) tasks is still a problem to be explored. Although there have been previous studies on the combination of SNNs and RL, most focus on robotic control problems with shallow networks or using the ANN-SNN conversion method to implement spiking deep Q networks (SDQN). In this study, we mathematically analyzed the problem of the disappearance of spiking signal features in SDQN and proposed a potential-based layer normalization (pbLN) method to train spiking deep Q networks directly. Experiment shows that compared with state-of-art ANN-SNN conversion method and other SDQN works, the proposed pbLN spiking deep Q networks (PL-SDQN) achieved better performance on Atari game tasks.

Training Spiking Neural Networks for Reinforcement Learning Tasks With Temporal Coding Method

Recent years witness an increasing demand for using spiking neural networks (SNNs) to implement artificial intelligent systems. There is a demand of combining SNNs with reinforcement learning architectures to find an effective training method. Recently, temporal coding method has been proposed to train spiking neural networks while preserving the asynchronous nature of spiking neurons to preserve the asynchronous nature of SNNs. We propose a training method that enables temporal coding method in RL tasks. To tackle the problem of high sparsity of spikes, we introduce a self-incremental variable to push each spiking neuron to fire, which makes SNNs fully differentiable. In addition, an encoding method is proposed to solve the problem of information loss of temporal-coded inputs. The experimental results show that the SNNs trained by our proposed method can achieve comparable performance of the state-of-the-art artificial neural networks in benchmark tasks of reinforcement learning.

Entropy-Aware Model Initialization for Effective Exploration in Deep Reinforcement Learning

Effective exploration is one of the critical factors affecting performance in deep reinforcement learning. Agents acquire data to learn the optimal policy through exploration, and if it is not guaranteed, the data quality deteriorates, which leads to performance degradation. This study investigates the effect of initial entropy, which significantly influences exploration, especially in the early learning stage. The results of this study on tasks with discrete action space show that (1) low initial entropy increases the probability of learning failure, (2) the distributions of initial entropy for various tasks are biased towards low values that inhibit exploration, and (3) the initial entropy for discrete action space varies with both the initial weight and task, making it hard to control. We then devise a simple yet powerful learning strategy to deal with these limitations, namely, entropy-aware model initialization. The proposed algorithm aims to provide a model with high initial entropy to a deep reinforcement learning algorithm for effective exploration. Our experiments showed that the devised learning strategy significantly reduces learning failures and enhances performance, stability, and learning speed.

Admiring the Great Mountain: A Celebration Special Issue in Honor of Stephen Grossberg's 80th Birthday

This editorial summarizes selected key contributions of Prof. Stephen Grossberg and describes the papers in this 80th birthday special issue in his honor. His productivity, creativity, and vision would each be enough to mark a scientist of the first caliber. In combination, they have resulted in contributions that have changed the entire discipline of neural networks. Grossberg has been tremendously influential in engineering, dynamical systems, and artificial intelligence as well. Indeed, he has been one of the most important mentors and role models in my career, and has done so with extraordinary generosity and encouragement. All authors in this special issue have taken great pleasure in hereby commemorating his extraordinary career and contributions.