scholarly journals Learning to Teach Reinforcement Learning Agents

2017 ◽  
Vol 1 (1) ◽  
pp. 21-42 ◽  
Author(s):  
Anestis Fachantidis ◽  
Matthew Taylor ◽  
Ioannis Vlahavas
Keyword(s):  
Reinforcement Learning ◽  
Learning Algorithm ◽  
Relevant Literature ◽  
Learning Approaches ◽  
Training Time ◽  
Factors Affecting ◽  
Learning Agents ◽  
Best Teachers ◽  
Previous Learning ◽  
Advice Quality

In this article, we study the transfer learning model of action advice under a budget. We focus on reinforcement learning teachers providing action advice to heterogeneous students playing the game of Pac-Man under a limited advice budget. First, we examine several critical factors affecting advice quality in this setting, such as the average performance of the teacher, its variance and the importance of reward discounting in advising. The experiments show that the best performers are not always the best teachers and reveal the non-trivial importance of the coefficient of variation (CV) as a statistic for choosing policies that generate advice. The CV statistic relates variance to the corresponding mean. Second, the article studies policy learning for distributing advice under a budget. Whereas most methods in the relevant literature rely on heuristics for advice distribution, we formulate the problem as a learning one and propose a novel reinforcement learning algorithm capable of learning when to advise or not. The proposed algorithm is able to advise even when it does not have knowledge of the student’s intended action and needs significantly less training time compared to previous learning approaches. Finally, in this article, we argue that learning to advise under a budget is an instance of a more generic learning problem: Constrained Exploitation Reinforcement Learning.

scholarly journals KoGuN: Accelerating Deep Reinforcement Learning via Integrating Human Suboptimal Knowledge

2020 ◽  
Author(s):  
Peng Zhang ◽  
Jianye Hao ◽  
Weixun Wang ◽  
Hongyao Tang ◽  
Yi Ma ◽  
...  
Keyword(s):  
Reinforcement Learning ◽  
Prior Knowledge ◽  
Learning Process ◽  
Learning Algorithm ◽  
Fuzzy Rule ◽  
Policy Network ◽  
Human Knowledge ◽  
Learning Agents ◽  
The Common ◽  
Low Performance

Reinforcement learning agents usually learn from scratch, which requires a large number of interactions with the environment. This is quite different from the learning process of human. When faced with a new task, human naturally have the common sense and use the prior knowledge to derive an initial policy and guide the learning process afterwards. Although the prior knowledge may be not fully applicable to the new task, the learning process is significantly sped up since the initial policy ensures a quick-start of learning and intermediate guidance allows to avoid unnecessary exploration. Taking this inspiration, we propose knowledge guided policy network (KoGuN), a novel framework that combines human prior suboptimal knowledge with reinforcement learning. Our framework consists of a fuzzy rule controller to represent human knowledge and a refine module to finetune suboptimal prior knowledge. The proposed framework is end-to-end and can be combined with existing policy-based reinforcement learning algorithm. We conduct experiments on several control tasks. The empirical results show that our approach, which combines suboptimal human knowledge and RL, achieves significant improvement on learning efficiency of flat RL algorithms, even with very low-performance human prior knowledge.

scholarly journals BUILDING AN ARTIFICIAL STOCK MARKET POPULATED BY REINFORCEMENT‐LEARNING AGENTS

2009 ◽  
Vol 10 (4) ◽  
pp. 329-341 ◽  
Author(s):  
Aleksandras Vytautas Rutkauskas ◽  
Tomas Ramanauskas
Keyword(s):  
Reinforcement Learning ◽  
Stock Market ◽  
Learning Algorithm ◽  
Self Regulation ◽  
Market Model ◽  
Emergent Properties ◽  
Q Learning ◽  
Evolutionary Selection ◽  
Learning Agents ◽  
Artificial Stock Market

In this paper we propose an artificial stock market model based on interaction of heterogeneous agents whose forward-looking behaviour is driven by the reinforcement-learning algorithm combined with some evolutionary selection mechanism. We use the model for the analysis of market self-regulation abilities, market efficiency and determinants of emergent properties of the financial market. Distinctive and novel features of the model include strong emphasis on the economic content of individual decision-making, application of the Q-learning algorithm for driving individual behaviour, and rich market setup. Along with that a parallel version of the model is presented, which is mainly based on research of current changes in the market, as well as on search of newly emerged consistent patterns, and which has been repeatedly used for optimal decisions’ search experiments in various capital markets.

scholarly journals Research on Tensor-Based Cooperative and Competitive in Multi-Agent Reinforcement Learning

2020 ◽  
Vol 4 (6) ◽  
Author(s):  
Tsega Weldu Araya ◽  
Md Rashed Ibn Nawab ◽  
A. P. Yuan Ling
Keyword(s):  
Reinforcement Learning ◽  
Learning Algorithm ◽  
Data Representation ◽  
Training Data ◽  
Two Dimensional ◽  
Multiple Agents ◽  
Learning Agents ◽  
Dimensional Array ◽  
Multi Agent ◽  
Agent Cooperation

As technology overgrows, the assortment of information and the density of work becomes demanding to manage. To resolve the density of employment and human labor, machine-learning (ML) technology developed. Reinforcement learning (RL) is the recent advancement of ML studies. Multi-agent reinforcement learning (MARL) is useful to train multiple agents in the surrounding environment. The previous research studies focused on two-agent cooperation. Their data representation was held in a two-dimensional array, which is called a matrix. The limitation of this two-dimensional array appears as the training data of agents increases. The growth in the training data of agents creates storage drawbacks and data redundancy. Our first aim in this research is to improve an algorithm that can represent MARL training in tensor. In MARL, multiple agents are work together to achieve joint work. To share the training records and data of numerous agents, we need to collect the previous cumulative experience of agents in tensor. Secondly, we will discover the agent's cooperation and competition, with local and global goals of agents in MARL. Local goals are the cooperation of agents in a group or team where we use the training model as a student and teacher agent. The global goal is the competition between two contrary teams to acquire the reward. All learning agents have their Q table for storing the individual agent's training data in an environment. The growth in the number of learning agents, their training experience in Q tables, and the requirement for representing multiple data become the most challenging issue. We introduce tensor to store various data to resolve the challenges for data representation in multiple agent associations. Tensor is expressed as the three-dimensional array, although it is an N-way array, which is useful for representing and accessing numerous data. Finally, we will implement an algorithm for learning three cooperative agents against the opposed team using a tensor-based framework in the Q learning algorithm. We will provide an algorithm that can store the training records and data of multiple agents. Tensor advances to get a small storage size than the matrix for the training records of agents. Although three agent cooperation benefits to having maximum optimal reward.

scholarly journals Application of Deep Q-learning and double Deep Q-learning algorithms to the task of control an inverted pendulum

2020 ◽  
pp. 7-25
Author(s):  
Alla Evseenko ◽  
◽  
Dmitrii Romannikov ◽  
Keyword(s):  
Artificial Intelligence ◽  
Machine Learning ◽  
Reinforcement Learning ◽  
Inverted Pendulum ◽  
Learning Algorithm ◽  
Learning Algorithms ◽  
Training Time ◽  
Q Learning ◽  
Research Areas ◽  
Wide Range

Today, such a branch of science as «artificial intelligence» is booming in the world. Systems built on the basis of artificial intelligence methods have the ability to perform functions that are traditionally considered the prerogative of man. Artificial intelligence has a wide range of research areas. One such area is machine learning. This article discusses the algorithms of one of the approaches of machine learning – reinforcement learning (RL), according to which a lot of research and development has been carried out over the past seven years. Development and research on this approach is mainly carried out to solve problems in Atari 2600 games or in other similar ones. In this article, reinforcement training will be applied to one of the dynamic objects – an inverted pendulum. As a model of this object, we consider a model of an inverted pendulum on a cart taken from the Gym library, which contains many models that are used to test and analyze reinforcement learning algorithms. The article describes the implementation and study of two algorithms from this approach, Deep Q-learning and Double Deep Q-learning. As a result, training, testing and training time graphs for each algorithm are presented, on the basis of which it is concluded that it is desirable to use the Double Deep Q-learning algorithm, because the training time is approximately 2 minutes and provides the best control for the model of an inverted pendulum on a cart.

scholarly journals Improving Maneuver Strategy in Air Combat by Alternate Freeze Games with a Deep Reinforcement Learning Algorithm

2020 ◽  
Vol 2020 ◽  
pp. 1-17
Author(s):  
Zhuang Wang ◽  
Hui Li ◽  
Haolin Wu ◽  
Zhaoxin Wu
Keyword(s):  
Reinforcement Learning ◽  
Learning Algorithm ◽  
Adaptive Strategy ◽  
Simulation Software ◽  
Learning Agents ◽  
Combat Simulation ◽  
Air Combat ◽  
Reward Shaping ◽  
Flight Level ◽  
The One

In a one-on-one air combat game, the opponent’s maneuver strategy is usually not deterministic, which leads us to consider a variety of opponent’s strategies when designing our maneuver strategy. In this paper, an alternate freeze game framework based on deep reinforcement learning is proposed to generate the maneuver strategy in an air combat pursuit. The maneuver strategy agents for aircraft guidance of both sides are designed in a flight level with fixed velocity and the one-on-one air combat scenario. Middleware which connects the agents and air combat simulation software is developed to provide a reinforcement learning environment for agent training. A reward shaping approach is used, by which the training speed is increased, and the performance of the generated trajectory is improved. Agents are trained by alternate freeze games with a deep reinforcement algorithm to deal with nonstationarity. A league system is adopted to avoid the red queen effect in the game where both sides implement adaptive strategies. Simulation results show that the proposed approach can be applied to maneuver guidance in air combat, and typical angle fight tactics can be learnt by the deep reinforcement learning agents. For the training of an opponent with the adaptive strategy, the winning rate can reach more than 50%, and the losing rate can be reduced to less than 15%. In a competition with all opponents, the winning rate of the strategic agent selected by the league system is more than 44%, and the probability of not losing is about 75%.

Weighted Densely Connected Convolutional Networks for Reinforcement Learning

2019 ◽  
Vol 34 (04) ◽  
pp. 2052001 ◽  
Author(s):  
Min Xia ◽  
Wenzhu Song ◽  
Xudong Sun ◽  
Jia Liu ◽  
Tao Ye ◽  
...  
Keyword(s):  
Reinforcement Learning ◽  
Learning Algorithm ◽  
Weight Coefficient ◽  
Cross Layer ◽  
Current Layer ◽  
Feature Maps ◽  
Training Time ◽  
Simulation Results ◽  
The Cross ◽  
Reinforcement Learning Algorithm

A weighted densely connected convolution network (W-DenseNet) is proposed for reinforcement learning in this work. The W-DenseNet can maximize the information flow between all layers in the network by cross layer connection, which can reduce the phenomenon of gradient vanishing and degradation, and greatly improves the speed of training convergence. The weight coefficient introduced in W-DenseNet, the current layer received all the previous layers’ feature maps with different initial weights, which can extract feature information of different layers more effectively according to tasks. According to the weight adjusted by learning, the cross-layer connection is pruned to remove the cross-layer connection with smaller weight, so as to reduce the number of cross-layer. In this work, GridWorld and FlappyBird games are used for simulation. The simulation results of deep reinforcement learning based on W-DenseNet are compared with the traditional deep reinforcement learning algorithm and reinforcement learning algorithm based on DenseNet. The simulation results show that the proposed W-DenseNet method can make the results more convergent, reduce the training time, and obtain more stable results.

A formal methods approach to interpretable reinforcement learning for robotic planning

2019 ◽  
Vol 4 (37) ◽  
pp. eaay6276 ◽  
Author(s):  
Xiao Li ◽  
Zachary Serlin ◽  
Guang Yang ◽  
Calin Belta
Keyword(s):  
Reinforcement Learning ◽  
Formal Methods ◽  
Learning Algorithm ◽  
A Priori ◽  
Generation Process ◽  
Learning Approaches ◽  
Learning Agent ◽  
Domain Specific Knowledge ◽  
Robotic Planning ◽  
High Level

Growing interest in reinforcement learning approaches to robotic planning and control raises concerns of predictability and safety of robot behaviors realized solely through learned control policies. In addition, formally defining reward functions for complex tasks is challenging, and faulty rewards are prone to exploitation by the learning agent. Here, we propose a formal methods approach to reinforcement learning that (i) provides a formal specification language that integrates high-level, rich, task specifications with a priori, domain-specific knowledge; (ii) makes the reward generation process easily interpretable; (iii) guides the policy generation process according to the specification; and (iv) guarantees the satisfaction of the (critical) safety component of the specification. The main ingredients of our computational framework are a predicate temporal logic specifically tailored for robotic tasks and an automaton-guided, safe reinforcement learning algorithm based on control barrier functions. Although the proposed framework is quite general, we motivate it and illustrate it experimentally for a robotic cooking task, in which two manipulators worked together to make hot dogs.

scholarly journals A Multiagent Reinforcement Learning Algorithm with Non-linear Dynamics

2008 ◽  
Vol 33 ◽  
pp. 521-549 ◽  
Author(s):  
S. Abdallah ◽  
V. Lesser
Keyword(s):  
Reinforcement Learning ◽  
Nash Equilibrium ◽  
Nash Equilibria ◽  
Learning Algorithm ◽  
A Priori ◽  
Linear Dynamics ◽  
Learning Agents ◽  
Multiagent Reinforcement Learning ◽  
Non Linear ◽  
Linear Nature

Several multiagent reinforcement learning (MARL) algorithms have been proposed to optimize agents' decisions. Due to the complexity of the problem, the majority of the previously developed MARL algorithms assumed agents either had some knowledge of the underlying game (such as Nash equilibria) and/or observed other agents actions and the rewards they received. We introduce a new MARL algorithm called the Weighted Policy Learner (WPL), which allows agents to reach a Nash Equilibrium (NE) in benchmark 2-player-2-action games with minimum knowledge. Using WPL, the only feedback an agent needs is its own local reward (the agent does not observe other agents actions or rewards). Furthermore, WPL does not assume that agents know the underlying game or the corresponding Nash Equilibrium a priori. We experimentally show that our algorithm converges in benchmark two-player-two-action games. We also show that our algorithm converges in the challenging Shapley's game where previous MARL algorithms failed to converge without knowing the underlying game or the NE. Furthermore, we show that WPL outperforms the state-of-the-art algorithms in a more realistic setting of 100 agents interacting and learning concurrently. An important aspect of understanding the behavior of a MARL algorithm is analyzing the dynamics of the algorithm: how the policies of multiple learning agents evolve over time as agents interact with one another. Such an analysis not only verifies whether agents using a given MARL algorithm will eventually converge, but also reveals the behavior of the MARL algorithm prior to convergence. We analyze our algorithm in two-player-two-action games and show that symbolically proving WPL's convergence is difficult, because of the non-linear nature of WPL's dynamics, unlike previous MARL algorithms that had either linear or piece-wise-linear dynamics. Instead, we numerically solve WPL's dynamics differential equations and compare the solution to the dynamics of previous MARL algorithms.

scholarly journals A Study on an Enhanced Autonomous Driving Simulation Model Based on Reinforcement Learning Using a Collision Prevention Model

Electronics ◽  
2021 ◽  
Vol 10 (18) ◽  
pp. 2271
Author(s):  
Jong-Hoon Kim ◽  
Jun-Ho Huh ◽  
Se-Hoon Jung ◽  
Chun-Bo Sim
Keyword(s):  
Reinforcement Learning ◽  
Performance Test ◽  
Learning Algorithm ◽  
Autonomous Driving ◽  
Urban Environments ◽  
Test Results ◽  
Learning Agents ◽  
Collision Prevention ◽  
Prevention Model ◽  
Definition Of

This paper set out to revise and improve existing autonomous driving models using reinforcement learning, thus proposing a reinforced autonomous driving prediction model. The paper conducted training for a reinforcement learning model using DQN, a reinforcement learning algorithm. The main aim of this paper was to reduce the time spent on training and improve self-driving performance. Rewards for reinforcement learning agents were developed to mimic human driving behavior as much as possible. High rewards were given for greater distance travelled within lanes and higher speed. Negative rewards were given when a vehicle crossed into other lanes or had a collision. Performance evaluation was carried out in urban environments without pedestrians. The performance test results show that the model with the collision prevention model exhibited faster performance improvement within the same time compared to when the model was not applied. However, vulnerabilities to factors such as pedestrians and vehicles approaching from the side were not addressed, and the lack of stability in the definition of compensation functions and limitations with respect to the excessive use of memory were shown.

scholarly journals Tactics of Adversarial Attack on Deep Reinforcement Learning Agents

2017 ◽  
Author(s):  
Yen-Chen Lin ◽  
Zhang-Wei Hong ◽  
Yuan-Hong Liao ◽  
Meng-Li Shih ◽  
Ming-Yu Liu ◽  
...  
Keyword(s):  
Reinforcement Learning ◽  
Learning Algorithm ◽  
Generative Model ◽  
Small Subset ◽  
Time Step ◽  
Learning Agents ◽  
Planning Algorithm ◽  
Adversarial Examples ◽  
Novel Method ◽  
Adversarial Example

We introduce two tactics, namely the strategically-timed attack and the enchanting attack, to attack reinforcement learning agents trained by deep reinforcement learning algorithms using adversarial examples. In the strategically-timed attack, the adversary aims at minimizing the agent's reward by only attacking the agent at a small subset of time steps in an episode. Limiting the attack activity to this subset helps prevent detection of the attack by the agent. We propose a novel method to determine when an adversarial example should be crafted and applied. In the enchanting attack, the adversary aims at luring the agent to a designated target state. This is achieved by combining a generative model and a planning algorithm: while the generative model predicts the future states, the planning algorithm generates a preferred sequence of actions for luring the agent. A sequence of adversarial examples is then crafted to lure the agent to take the preferred sequence of actions. We apply the proposed tactics to the agents trained by the state-of-the-art deep reinforcement learning algorithm including DQN and A3C. In 5 Atari games, our strategically-timed attack reduces as much reward as the uniform attack (i.e., attacking at every time step) does by attacking the agent 4 times less often. Our enchanting attack lures the agent toward designated target states with a more than 70% success rate. Example videos are available at http://yclin.me/adversarial_attack_RL/.

Sign in / Sign up

Export Citation Format

Share Document