Natural disasters can cause substantial negative socio-economic impacts around the world, due to mortality, relocation, rates, and reconstruction decisions. Robotics has been successfully applied to identify and rescue victims during the occurrence of a natural hazard. However, little effort has been taken to deploy solutions where an autonomous robot can save the life of a citizen by itself relocating it, without the need to wait for a rescue team composed of humans. Reinforcement learning approaches can be used to deploy such a solution, however, one of the most famous algorithms to deploy it, the Q-learning, suffers from biased results generated when performing its learning routines. In this research a solution for citizen relocation based on Partially Observable Markov Decision Processes is adopted, where the capability of the Double Q-learning in relocating citizens during a natural hazard is evaluated under a proposed hazard simulation engine based on a grid world. The performance of the solution was measured as a success rate of a citizen relocation procedure, where the results show that the technique portrays a performance above 100% for easy scenarios and near 50% for hard ones.
Most reinforcement learning algorithms implicitly assume strong synchrony. We present novel attacks targeting Q-learning that exploit a vulnerability entailed by this assumption by delaying the reward signal for a limited time period. We consider two types of attack goals: targeted attacks, which aim to cause a target policy to be learned, and untargeted attacks, which simply aim to induce a policy with a low reward. We evaluate the efficacy of the proposed attacks through a series of experiments. Our first observation is that reward-delay attacks are extremely effective when the goal is simply to minimize reward. Indeed, we find that even naive baseline reward-delay attacks are also highly successful in minimizing the reward. Targeted attacks, on the other hand, are more challenging, although we nevertheless demonstrate that the proposed approaches remain highly effective at achieving the attacker's targets. In addition, we introduce a second threat model that captures a minimal mitigation that ensures that rewards cannot be used out of sequence. We find that this mitigation remains insufficient to ensure robustness to attacks that delay, but preserve the order, of rewards.
In reinforcement learning, Monte Carlo algorithms update the Q function by averaging the episodic returns. In the Monte Carlo UCB (MC-UCB) algorithm, the action taken in each state is the action that maximizes the Q function plus a UCB exploration term, which biases the choice of actions to those that have been chosen less frequently. Although there has been significant work on establishing regret bounds for MC-UCB, most of that work has been focused on finite-horizon versions of the problem, for which each episode terminates after a constant number of steps. For such finite-horizon problems, the optimal policy depends both on the current state and the time within the episode. However, for many natural episodic problems, such as games like Go and Chess and robotic tasks, the episode is of random length and the optimal policy is stationary. For such environments, it is an open question whether the Q-function in MC-UCB will converge to the optimal Q function; we conjecture that, unlike Q-learning, it does not converge for all MDPs. We nevertheless show that for a large class of MDPs, which includes stochastic MDPs such as blackjack and deterministic MDPs such as Go, the Q-function in MC-UCB converges almost surely to the optimal Q function. An immediate corollary of this result is that it also converges almost surely for all finite-horizon MDPs. We also provide numerical experiments, providing further insights into MC-UCB.
We consider the problem of sequential recommendations, where at each step an agent proposes some slate of $N$ distinct items to a user from a much larger catalog of size $K>>N$. The user has unknown preferences towards the recommendations and the agent takes sequential actions that optimise (in our case minimise) some user-related cost, with the help of Reinforcement Learning. The possible item combinations for a slate is $\binom{K}{N}$, an enormous number rendering value iteration methods intractable. We prove that the slate-MDP can actually be decomposed using just $K$ item-related $Q$ functions per state, which describe the problem in a more compact and efficient way. Based on this, we propose a novel model-free SARSA and Q-learning algorithm that performs $N$ parallel iterations per step, without any prior user knowledge. We call this method \texttt{SlateFree}, i.e. free-of-slates, and we show numerically that it converges very fast to the exact optimum for arbitrary user profiles, and that it outperforms alternatives from the literature.
Board games, with the exception of solo games, need at least one other player to play. Because of this, we created Artificial Intelligent (AI) agents to play against us when an opponent is missing. These AI agents are created in a number of ways, but one challenge with these agents is that an agent can have superior ability compared to us. In this work, we describe how to create weaker AI agents that play board games. We use Tic-Tac-Toe, Nine-Men's Morris, and Mancala, and our technique uses a Reinforcement Learning model where an agent uses the Q-learning algorithm to learn these games. We show how these agents can learn to play the board game perfectly, and we then describe our approach to making weaker versions of these agents. Finally, we provide a methodology to compare AI agents.
This paper investigates the problem of bounding possible output from a counterfactual query given a set of observational data. While various works of literature have described methodologies to generate efficient algorithms that provide an optimal bound for the counterfactual query, all of them assume a finite-horizon causal diagram. This paper aims to extend the previous work by modifying Q-learning algorithm to provide informative bounds of a causal query given an infinite-horizon causal diagram. Through simulations, our algorithms are proven to perform better compared to existing algorithm.
We present a novel direct data-driven algorithm that learns an optimal control policy for the Bilinear Biquadratic Regulator (BBR) for an unknown bilinear system. The BBR is difficult to solve owing to the presence of the nonlinear biquadratic performance index and the bilinear cross-term in the dynamics. To address these difficulties, we apply several transformations on the state decision variables to obtain a nonlinear optimization problem with a linear performance index and affine (in the parameterized control) state-dependent equality. The adroit use of the Hamiltonian and Pontryagin's Minimum Principle allows us to derive a pair of first-order necessary conditions that, at each point in time, are easily solvable linear matrix equalities (LMEs) which give the optimal state-dependent control law. We then use the marginal sample autocorrelation of the collected data to obtain a direct data-driven equivalent of these LMEs. We demonstrate the performance of the proposed algorithm via illustrative numerical examples.
Many applications of reinforcement learning can be formalized as goal-conditioned environments, where, in each episode, there is a "goal" that affects the rewards obtained during that episode but does not affect the dynamics. Various techniques have been proposed to improve performance in goal-conditioned environments, such as automatic curriculum generation and goal relabeling. In this work, we explore a connection between off-policy reinforcement learning in goal-conditioned settings and knowledge distillation. In particular: the current Q-value function and the target Q-value estimate are both functions of the goal, and we would like to train the Q-value function to match its target for all goals. We therefore apply Gradient-Based Attention Transfer (Zagoruyko and Komodakis 2017), a knowledge distillation technique, to the Q-function update. We empirically show that this can improve the performance of goal-conditioned off-policy reinforcement learning when the space of goals is high-dimensional. We also show that this technique can be adapted to allow for efficient learning in the case of multiple simultaneous sparse goals, where the agent can attain a reward by achieving any one of a large set of objectives, all specified at test time. Finally, to provide theoretical support, we give examples of classes of environments where (under some assumptions) standard off-policy algorithms such as DDPG require at least O(d^2) replay buffer transitions to learn an optimal policy, while our proposed technique requires only O(d) transitions, where d is the dimensionality of the goal and state space. Code is available at https://github.com/alevine0/ReenGAGE.
Object Goal Navigation requires a robot to find and navigate to an instance of a target object class in a previously unseen environment. Our framework incrementally builds a semantic map of the environment over time, and then repeatedly selects a long-term goal ('where to go') based on the semantic map to locate the target object instance. Long-term goal selection is formulated as a vision-based deep reinforcement learning problem. Specifically, an Encoder Network is trained to extract high-level features from a semantic map and select a long-term goal. In addition, we incorporate data augmentation and Q-function regularization to make the long-term goal selection more effective. We report experimental results using the photo-realistic Gibson benchmark dataset in the AI Habitat 3D simulation environment to demonstrate substantial performance improvement on standard measures in comparison with a state of the art data-driven baseline.
The widespread adoption of distributed energy resources, and the advent of smart grid technologies, have allowed traditionally passive power system users to become actively involved in energy trading. Recognizing the fact that the traditional centralized grid-driven energy markets offer minimal profitability to these users, recent research has shifted focus towards decentralized peer-to-peer (P2P) energy markets. In these markets, users trade energy with each other, with higher benefits than buying or selling to the grid. However, most researches in P2P energy trading largely overlook the user perception in the trading process, assuming constant availability, participation, and full compliance. As a result, these approaches may result in negative attitudes and reduced engagement over time. In this paper, we design an automated P2P energy market that takes user perception into account. We employ prospect theory to model the user perception and formulate an optimization framework to maximize the buyer's perception while matching demand and production. Given the non-linear and non-convex nature of the optimization problem, we propose Differential Evolution-based Algorithm for Trading Energy called DEbATE. Additionally, we introduce a risk-sensitive Q-learning algorithm, named Pricing mechanism with Q-learning and Risk-sensitivity (PQR), which learns the optimal price for sellers considering their perceived utility. Results based on real traces of energy consumption and production, as well as realistic prospect theory functions, show that our approach achieves a 26% higher perceived value for buyers and generates 7% more reward for sellers, compared to a recent state of the art approach.
Conventional anti-jamming methods mainly focus on preventing single jammer attacks with an invariant jamming policy or jamming attacks from multiple jammers with similar jamming policies. These anti-jamming methods are ineffective against a single jammer following several different jamming policies or multiple jammers with distinct policies. Therefore, this paper proposes an anti-jamming method that can adapt its policy to the current jamming attack. Moreover, for the multiple jammers scenario, an anti-jamming method that estimates the future occupied channels using the jammers' occupied channels in previous time slots is proposed. In both single and multiple jammers scenarios, the interaction between the users and jammers is modeled using recurrent neural networks (RNN)s. The performance of the proposed anti-jamming methods is evaluated by calculating the users' successful transmission rate (STR) and ergodic rate (ER), and compared to a baseline based on Q-learning (DQL). Simulation results show that for the single jammer scenario, all the considered jamming policies are perfectly detected and high STR and ER are maintained. Moreover, when 70 % of the spectrum is under jamming attacks from multiple jammers, the proposed method achieves an STR and ER greater than 75 % and 80 %, respectively. These values rise to 90 % when 30 % of the spectrum is under jamming attacks. In addition, the proposed anti-jamming methods significantly outperform the DQL method for all the considered cases and jamming scenarios.
Offline reinforcement learning (RL), which aims to learn an optimal policy using a previously collected static dataset, is an important paradigm of RL. Standard RL methods often perform poorly in this regime due to the function approximation errors on out-of-distribution actions. While a variety of regularization methods have been proposed to mitigate this issue, they are often constrained by policy classes with limited expressiveness that can lead to highly suboptimal solutions. In this paper, we propose representing the policy as a diffusion model, a recent class of highly-expressive deep generative models. We introduce Diffusion Q-learning (Diffusion-QL) that utilizes a conditional diffusion model to represent the policy. In our approach, we learn an action-value function and we add a term maximizing action-values into the training loss of the conditional diffusion model, which results in a loss that seeks optimal actions that are near the behavior policy. We show the expressiveness of the diffusion model-based policy, and the coupling of the behavior cloning and policy improvement under the diffusion model both contribute to the outstanding performance of Diffusion-QL. We illustrate the superiority of our method compared to prior works in a simple 2D bandit example with a multimodal behavior policy. We then show that our method can achieve state-of-the-art performance on the majority of the D4RL benchmark tasks.
This work proposes a novel resource allocation strategy for anti-jamming in Cognitive Radio using Active Inference ($\textit{AIn}$), and a cognitive-UAV is employed as a case study. An Active Generalized Dynamic Bayesian Network (Active-GDBN) is proposed to represent the external environment that jointly encodes the physical signal dynamics and the dynamic interaction between UAV and jammer in the spectrum. We cast the action and planning as a Bayesian inference problem that can be solved by avoiding surprising states (minimizing abnormality) during online learning. Simulation results verify the effectiveness of the proposed $\textit{AIn}$ approach in minimizing abnormalities (maximizing rewards) and has a high convergence speed by comparing it with the conventional Frequency Hopping and Q-learning.
We propose a compositional approach to synthesize policies for networks of continuous-space stochastic control systems with unknown dynamics using model-free reinforcement learning (RL). The approach is based on implicitly abstracting each subsystem in the network with a finite Markov decision process with unknown transition probabilities, synthesizing a strategy for each abstract model in an assume-guarantee fashion using RL, and then mapping the results back over the original network with approximate optimality guarantees. We provide lower bounds on the satisfaction probability of the overall network based on those over individual subsystems. A key contribution is to leverage the convergence results for adversarial RL (minimax Q-learning) on finite stochastic arenas to provide control strategies maximizing the probability of satisfaction over the network of continuous-space systems. We consider finite-horizon properties expressed in the syntactically co-safe fragment of linear temporal logic. These properties can readily be converted into automata-based reward functions, providing scalar reward signals suitable for RL. Since such reward functions are often sparse, we supply a potential-based reward shaping technique to accelerate learning by producing dense rewards. The effectiveness of the proposed approaches is demonstrated via two physical benchmarks including regulation of a room temperature network and control of a road traffic network.
Vehicle-to-Infrastructure (V2I) communication is becoming critical for the enhanced reliability of autonomous vehicles (AVs). However, the uncertainties in the road-traffic and AVs' wireless connections can severely impair timely decision-making. It is thus critical to simultaneously optimize the AVs' network selection and driving policies in order to minimize road collisions while maximizing the communication data rates. In this paper, we develop a reinforcement learning (RL) framework to characterize efficient network selection and autonomous driving policies in a multi-band vehicular network (VNet) operating on conventional sub-6GHz spectrum and Terahertz (THz) frequencies. The proposed framework is designed to (i) maximize the traffic flow and minimize collisions by controlling the vehicle's motion dynamics (i.e., speed and acceleration) from autonomous driving perspective, and (ii) maximize the data rates and minimize handoffs by jointly controlling the vehicle's motion dynamics and network selection from telecommunication perspective. We cast this problem as a Markov Decision Process (MDP) and develop a deep Q-learning based solution to optimize the actions such as acceleration, deceleration, lane-changes, and AV-base station assignments for a given AV's state. The AV's state is defined based on the velocities and communication channel states of AVs. Numerical results demonstrate interesting insights related to the inter-dependency of vehicle's motion dynamics, handoffs, and the communication data rate. The proposed policies enable AVs to adopt safe driving behaviors with improved connectivity.
Task scheduling is a critical problem when one user offloads multiple different tasks to the edge server. When a user has multiple tasks to offload and only one task can be transmitted to server at a time, while server processes tasks according to the transmission order, the problem is NP-hard. However, it is difficult for traditional optimization methods to quickly obtain the optimal solution, while approaches based on reinforcement learning face with the challenge of excessively large action space and slow convergence. In this paper, we propose a Digital Twin (DT)-assisted RL-based task scheduling method in order to improve the performance and convergence of the RL. We use DT to simulate the results of different decisions made by the agent, so that one agent can try multiple actions at a time, or, similarly, multiple agents can interact with environment in parallel in DT. In this way, the exploration efficiency of RL can be significantly improved via DT, and thus RL can converges faster and local optimality is less likely to happen. Particularly, two algorithms are designed to made task scheduling decisions, i.e., DT-assisted asynchronous Q-learning (DTAQL) and DT-assisted exploring Q-learning (DTEQL). Simulation results show that both algorithms significantly improve the convergence speed of Q-learning by increasing the exploration efficiency.
Cost-effective asset management is an area of interest across several industries. Specifically, this paper develops a deep reinforcement learning (DRL) solution to automatically determine an optimal rehabilitation policy for continuously deteriorating water pipes. We approach the problem of rehabilitation planning in an online and offline DRL setting. In online DRL, the agent interacts with a simulated environment of multiple pipes with distinct lengths, materials, and failure rate characteristics. We train the agent using deep Q-learning (DQN) to learn an optimal policy with minimal average costs and reduced failure probability. In offline learning, the agent uses static data, e.g., DQN replay data, to learn an optimal policy via a conservative Q-learning algorithm without further interactions with the environment. We demonstrate that DRL-based policies improve over standard preventive, corrective, and greedy planning alternatives. Additionally, learning from the fixed DQN replay dataset in an offline setting further improves the performance. The results warrant that the existing deterioration profiles of water pipes consisting of large and diverse states and action trajectories provide a valuable avenue to learn rehabilitation policies in the offline setting, which can be further fine-tuned using the simulator.
Compared to on-policy counterparts, off-policy model-free deep reinforcement learning can improve data efficiency by repeatedly using the previously gathered data. However, off-policy learning becomes challenging when the discrepancy between the underlying distributions of the agent's policy and collected data increases. Although the well-studied importance sampling and off-policy policy gradient techniques were proposed to compensate for this discrepancy, they usually require a collection of long trajectories and induce additional problems such as vanishing/exploding gradients or discarding many useful experiences, which eventually increases the computational complexity. Moreover, their generalization to either continuous action domains or policies approximated by deterministic deep neural networks is strictly limited. To overcome these limitations, we introduce a novel policy similarity measure to mitigate the effects of such discrepancy in continuous control. Our method offers an adequate single-step off-policy correction that is applicable to deterministic policy networks. Theoretical and empirical studies demonstrate that it can achieve a "safe" off-policy learning and substantially improve the state-of-the-art by attaining higher returns in fewer steps than the competing methods through an effective schedule of the learning rate in Q-learning and policy optimization.
Transfer learning is an increasingly common approach for developing performant RL agents. However, it is not well understood how to define the relationship between the source and target tasks, and how this relationship contributes to successful transfer. We present an algorithm called Structural Similarity for Two MDPS, or SS2, that calculates a state similarity measure for states in two finite MDPs based on previously developed bisimulation metrics, and show that the measure satisfies properties of a distance metric. Then, through empirical results with GridWorld navigation tasks, we provide evidence that the distance measure can be used to improve transfer performance for Q-Learning agents over previous implementations.
For over a decade now, robotics and the use of artificial agents have become a common thing.Testing the performance of new path finding or search space optimization algorithms has also become a challenge as they require simulation or an environment to test them.The creation of artificial environments with artificial agents is one of the methods employed to test such algorithms.Games have also become an environment to test them.The performance of the algorithms can be compared by using artificial agents that will behave according to the algorithm in the environment they are put in.The performance parameters can be, how quickly the agent is able to differentiate between rewarding actions and hostile actions.This can be tested by placing the agent in an environment with different types of hurdles and the goal of the agent is to reach the farthest by taking decisions on actions that will lead to avoiding all the obstacles.The environment chosen is a game called "Flappy Bird".The goal of the game is to make the bird fly through a set of pipes of random heights.The bird must go in between these pipes and must not hit the top, the bottom, or the pipes themselves.The actions that the bird can take are either to flap its wings or drop down with gravity.The algorithms that are enforced on the artificial agents are NeuroEvolution of Augmenting Topologies (NEAT) and Reinforcement Learning.The NEAT algorithm takes an "N" initial population of artificial agents.They follow genetic algorithms by considering an objective function, crossover, mutation, and augmenting topologies.Reinforcement learning, on the other hand, remembers the state, the action taken at that state, and the reward received for the action taken using a single agent and a Deep Q-learning Network.The performance of the NEAT algorithm improves as the initial population of the artificial agents is increased.