Reinforcement Learning (RL)

Ape-X DQN is a variant of a DQN with some components of Rainbow-DQN that utilizes distributed prioritized experience replay through the Ape-X architecture.

Gated Transformer-XL, or GTrXL, is a Transformer-based architecture for reinforcement learning. It introduces architectural modifications that improve the stability and learning speed of the original Transformer and XL variant. Changes include: - Placing the layer normalization on only the input stream of the submodules. A key benefit to this reordering is that it now enables an identity map from the input of the transformer at the first layer to the output of the transformer after the last layer. This is in contrast to the canonical transformer, where there are a series of layer normalization operations that non-linearly transform the state encoding. - Replacing residual connections with gating layers. The authors' experiments found that GRUs were the most effective form of gating.

GridMask is a data augmentation method that randomly removes some pixels of an input image. Unlike other methods, the region that the algorithm removes is neither a continuous region nor random pixels in dropout. Instead, the algorithm removes a region with disconnected pixel sets, as shown in the Figure. We express the setting as $$ \tilde{\mathbf{x}}=\mathbf{x} \times M $$ where $\mathbf{x} \in R^{H \times W \times C}$ represents the input image, $M \in$ $\{0,1\}^{H \times W}$ is the binary mask that stores pixels to be removed, and $\tilde{\mathbf{x}} \in R^{H \times W \times C}$ is the result produced by the algorithm. For the binary mask $M$, if $M_{i, j}=1$ we keep pixel $(i, j)$ in the input image; otherwise we remove it. GridMask is applied after the image normalization operation. The shape of $M$ looks like a grid, as shown in the Figure . Four numbers $\left(r, d, \delta_{x}, \delta_{y}\right)$ are used to represent a unique $M$. Every mask is formed by tiling the units. $r$ is the ratio of the shorter gray edge in a unit. $d$ is the length of one unit. $\delta\_{x}$ and $\delta\_{y}$ are the distances between the first intact unit and boundary of the image.

ACKTR, or Actor Critic with Kronecker-factored Trust Region, is an actor-critic method for reinforcement learning that applies trust region optimization using a recently proposed Kronecker-factored approximation to the curvature. The method extends the framework of natural policy gradient and optimizes both the actor and the critic using Kronecker-factored approximate curvature (K-FAC) with trust region.

Non-parametric approximation of Q-values by storing all visited states and doing inference through k-Nearest Neighbors.

NoisyNet-DQN is a modification of a DQN that utilises noisy linear layers for exploration instead of $\epsilon$-greedy exploration as in the original DQN formulation.

Policy Similarity Metric, or PSM, is a similarity metric for measuring behavioral similarity between states in reinforcement learning. It assigns high similarity to states for which the optimal policies in those states as well as in future states are similar. PSM is reward-agnostic, making it more robust for generalization compared to approaches that rely on reward information.

TD-VAE, or Temporal Difference VAE, is a generative sequence model that learns representations containing explicit beliefs about states several steps into the future, and that can be rolled out directly without single-step transitions. TD-VAE is trained on pairs of temporally separated time points, using an analogue of temporal difference learning used in reinforcement learning.

TorchBeast is a platform for reinforcement learning (RL) research in PyTorch. It implements a version of the popular IMPALA algorithm for fast, asynchronous, parallel training of RL agents.

Ape-X DPG combines DDPG with distributed prioritized experience replay through the Ape-X architecture.

Blue River Controls is a tool that allows users to train and test reinforcement learning algorithms on real-world hardware. It features a simple interface based on OpenAI Gym, that works directly on both simulation and hardware.

Contrastive BERT is a reinforcement learning agent that combines a new contrastive loss and a hybrid LSTM-transformer architecture to tackle the challenge of improving data efficiency for RL. It uses bidirectional masked prediction in combination with a generalization of recent contrastive methods to learn better representations for transformers in RL, without the need of hand engineered data augmentations. For the architecture, a residual network is used to encode observations into embeddings $Y\_{t}$. $Y_{t}$ is fed through a causally masked GTrXL transformer, which computes the predicted masked inputs $X\_{t}$ and passes those together with $Y\_{t}$ to a learnt gate. The output of the gate is passed through a single LSTM layer to produce the values that we use for computing the RL loss. A contrastive loss is computed using predicted masked inputs $X_{t}$ and $Y_{t}$ as targets. For this, we do not use the causal mask of the Transformer.

FORK, or Forward Looking Actor is a type of actor for actor-critic algorithms. In particular, FORK includes a neural network that forecasts the next state given the current state and current action, called system network; and a neural network that forecasts the reward given a (state, action) pair, called reward network. With the system network and reward network, FORK can forecast the next state and consider the value of the next state when improving the policy.

Kalman Optimization for Value Approximation, or KOVA is a general framework for addressing uncertainties while approximating value-based functions in deep RL domains. KOVA minimizes a regularized objective function that concerns both parameter and noisy return uncertainties. It is feasible when using non-linear approximation functions as DNNs and can estimate the value in both on-policy and off-policy settings. It can be incorporated as a policy evaluation component in policy optimization algorithms.

MushroomRL is an open-source Python library developed to simplify the process of implementing and running Reinforcement Learning (RL) experiments. The architecture of MushroomRL is built in such a way that every component of an RL problem is already provided, and most of the time users can only focus on the implementation of their own algorithms and experiments. MushroomRL comes with a strongly modular architecture that makes it easy to understand how each component is structured and how it interacts with other ones; moreover it provides an exhaustive list of RL methodologies, such as:

NoisyNet-A3C is a modification of A3C that utilises noisy linear layers for exploration instead of $\epsilon$-greedy exploration as in the original DQN formulation.

NoisyNet-Dueling is a modification of a Dueling Network that utilises noisy linear layers for exploration instead of $\epsilon$-greedy exploration as in the original Dueling formulation.

Path Planning and Motion Control, or PPMC RL, is a training algorithm that teaches path planning and motion control to robots using reinforcement learning in a simulated environment. The focus is on promoting generalization where there are environmental uncertainties such as rough environments like lunar services. The algorithm is coupled with any generic reinforcement learning algorithm to teach robots how to respond to user commands and to travel to designated locations on a single neural network. The algorithm works independently of the robot structure, demonstrating that it works on a wheeled rover in addition to the past results on a quadruped walking robot.

Protagonist Antagonist Induced Regret Environment Design, or PAIRED, is an adversarial method for approximate minimax regret to generate environments for reinforcement learning. It introduces an antagonist which is allied with the environment generating adversary. The primary agent we are trying to train is the protagonist. The environment adversary’s goal is to design environments in which the antagonist achieves high reward and the protagonist receives low reward. If the adversary generates unsolvable environments, the antagonist and protagonist would perform the same and the adversary would get a score of zero, but if the adversary finds environments the antagonist solves and the protagonist does not solve, the adversary achieves a positive score. Thus, the environment adversary is incentivized to create challenging but feasible environments, in which the antagonist can outperform the protagonist. Moreover, as the protagonist learns to solves the simple environments, the antagonist must generate more complex environments to make the protagonist fail, increasing the complexity of the generated tasks and leading to automatic curriculum generation.

Robust Predictable Control, or RPC, is an RL algorithm for learning policies that uses only a few bits of information. RPC brings together ideas from information bottlenecks, model-based RL, and bits-back coding. The main idea of RPC is that if the agent can accurately predict the future, then the agent will not need to observe as many bits from future observations. Precisely, the agent will learn a latent dynamics model that predicts the next representation using the current representation and action. In addition to predicting the future, the agent can also decrease the number of bits by changing its behavior. States where the dynamics are hard to predict will require more bits, so the agent will prefer visiting states where its learned model can accurately predict the next state.

TayPO, or Taylor Expansion Policy Optimization, refers to a set of algorithms that apply the $k$-th order Taylor expansions for policy optimization. This generalizes prior work, including TRPO as a special case. It can be thought of unifying ideas from trust-region policy optimization and off-policy corrections. Taylor expansions share high-level similarities with both trust region policy search and off-policy corrections. To get high-level intuitions of such similarities, consider a simple 1D example of Taylor expansions. Given a sufficiently smooth real-valued function on the real line $f : \mathbb{R} \rightarrow \mathbb{R}$, the $k$-th order Taylor expansion of $f\left(x\right)$ at $x\_{0}$ is $$f\_{k}\left(x\right) = f\left(x\_{0}\right)+\sum^{k}\_{i=1}\left[f^{(i)}\left(x\_{0}\right)/i!\right]\left(x−x\_{0}\right)^{i}$$ where $f^{(i)}\left(x\_{0}\right)$ are the $i$-th order derivatives at $x\_{0}$. First, a common feature shared by Taylor expansions and trust-region policy search is the inherent notion of a trust region constraint. Indeed, in order for convergence to take place, a trust-region constraint is required $|x − x\_{0}| < R\left(f, x\_{0}\right)^{1}$. Second, when using the truncation as an approximation to the original function $f\_{K}\left(x\right) \approx f\left(x\right)$, Taylor expansions satisfy the requirement of off-policy evaluations: evaluate target policy with behavior data. Indeed, to evaluate the truncation $f\_{K}\left(x\right)$ at any $x$ (target policy), we only require the behavior policy "data" at $x\_{0}$ (i.e., derivatives $f^{(i)}\left(x\_{0}\right)$).

Generalized State-Dependent Exploration, or gSDE, is an exploration method for reinforcement learning that uses more general features and re-sampling the noise periodically. State-Dependent Exploration (SDE) is an intermediate solution for exploration that consists in adding noise as a function of the state $s\_{t}$, to the deterministic action $\mu\left(\mathbf{s}\_{t}\right)$. At the beginning of an episode, the parameters $\theta\_{\epsilon}$ of that exploration function are drawn from a Gaussian distribution. The resulting action $\mathbf{a}\_{t}$ is as follows: $$ \mathbf{a}\_{t}=\mu\left(\mathbf{s}\_{t} ; \theta\_{\mu}\right)+\epsilon\left(\mathbf{s}\_{t} ; \theta\_{\epsilon}\right), \quad \theta\_{\epsilon} \sim \mathcal{N}\left(0, \sigma^{2}\right) $$ This episode-based exploration is smoother and more consistent than the unstructured step-based exploration. Thus, during one episode, instead of oscillating around a mean value, the action a for a given state $s$ will be the same. In the case of a linear exploration function $\epsilon\left(\mathbf{s} ; \theta\_{\epsilon}\right)=\theta\_{\epsilon} \mathbf{s}$, by operation on Gaussian distributions, Rückstieß et al. show that the action element $\mathbf{a}\_{j}$ is normally distributed: $$ \pi]_{j}\left(\mathbf{a}\_{j} \mid \mathbf{s}\right) \sim \mathcal{N}\left(\mu\_{j}(\mathbf{s}), \hat{\sigma\_{j}}^{2}\right) $$ where $\hat{\sigma}$ is a diagonal matrix with elements $\hat{\sigma}\_{j}=\sqrt{\sum\_{i}\left(\sigma\_{i j} \mathbf{s}\_{i}\right)^{2}}$. Because we know the policy distribution, we can obtain the derivative of the log-likelihood $\log \pi(\mathbf{a} \mid \mathbf{s})$ with respect to the variance $\sigma$ : $$ \frac{\partial \log \pi(\mathbf{a} \mid \mathbf{s})}{\partial \sigma_{i j}}=\frac{\left(\mathbf{a}\_{j}-\mu\_{j}\right)^{2}-\hat{\sigma\_{j}}^{2}}{\hat{\sigma}\_{j}^{3}} \frac{\mathbf{s}\_{i}^{2} \sigma\_{i j}}{\hat{\sigma_{j}}} $$ This can be easily plugged into the likelihood ratio gradient estimator, which allows to adapt $\sigma$ during training. SDE is therefore compatible with standard policy gradient methods, while addressing most shortcomings of the unstructured exploration. For gSDE, two improvements are suggested: 1. We sample the parameters $\theta\_{\epsilon}$ of the exploration function every $n$ steps instead of every episode. 2. Instead of the state s, we can in fact use any features. We chose policy features $\mathbf{z}\_{\mu}\left(\mathbf{s} ; \theta\_{\mathbf{z}\_{\mu}}\right)$ (last layer before the deterministic output $\left.\mu(\mathbf{s})=\theta\_{\mu} \mathbf{z}\_{\mu}\left(\mathbf{s} ; \theta_{\mathbf{z}\_{\mu}}\right)\right)$ as input to the noise function $\epsilon\left(\mathbf{s} ; \theta\_{\epsilon}\right)=\theta\_{\epsilon} \mathbf{z}\_{\mu}(\mathbf{s})$

Stefano Guarnieri, Francesco Piazza, and Aurelio Uncini "Multilayer Feedforward Networks with Adaptive Spline Activation Function," IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 10, NO. 3, MAY 1999 Abstract — In this paper, a new adaptive spline activation function neural network (ASNN) is presented. Due to the ASNN’s high representation capabilities, networks with a small number of interconnections can be trained to solve both pattern recognition and data processing real-time problems. The main idea is to use a Catmull–Rom cubic spline as the neuron’s activation function, which ensures a simple structure suitable for both software and hardware implementation. Experimental results demonstrate improvements in terms of generalization capability and of learning speed in both pattern recognition and data processing tasks. Index Terms— Adaptive activation functions, function shape autotuning, generalization, generalized sigmoidal functions, multilayer perceptron, neural networks, spline neural networks.