Deep learning, reinforcement learning, and world models

Matsuo, Yutaka; LeCun, Yann; Sahani, Maneesh; Precup, Doina; Silver, David; Sugiyama, Masashi; Uchibe, Eiji; Morimoto, Jun

doi:10.1016/j.neunet.2022.03.037

Cited by 223 publications

(91 citation statements)

References 51 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…e essence of oral English pronunciation recognition based on machine learning [12][13][14] is to classify and recognize each word in the input audio. e essence of oral English pronunciation recognition based on deep learning [15][16][17] is the same. e difference is that the process is different.…”

Section: Introductionmentioning

confidence: 99%

A Deep Learning-Based Assisted Teaching System for Oral English

Liu

2022

Security and Communication Networks

View full text Add to dashboard Cite

The progress of global economic integration has forced English learners to have an urgent need to improve their oral English. College students’ oral English ability is currently the worst of the four abilities of listening, speaking, reading, and writing. The main reasons are internal and external. The internal reason is that the pronunciation characteristics of Chinese students are different from those of English. The external cause is that the practice environment and tools of oral English are not ideal, which affects the improvement of learners’ oral English. This study proposes using a deep learning algorithm (DLA) English in the evaluation of oral English quality to improve learners’ oral English level. The quality of oral English can be comprehensively evaluated in terms of pitch, speed of sound, and rhythm. The standard of pronunciation is the foundation of oral English and is the most critical factor. In many DLAs, the input unit of DNN at a certain moment and its upper and lower moment input units have no relationship and are independent of each other, and the timing dependencies of adjacent units are not fully considered. The results are generally not very good on speech recognition tasks. This study proposes a time-delay neural network (TDNN) and a long short-term memory (LSTM) network to calculate the posterior probability of the model state to model context-dependent features in order to solve this problem. The fusion model TDNN-LSTM is applied in the English spoken pronunciation recognition task. To compare the accuracy of oral English pronunciation, several classic DLAs are introduced. The experimental results show that the method described in this study has a number of advantages. Although the performance improvement of this method in terms of recognition accuracy is not large, a certain degree of improvement is also very important for the oral English teaching assistant system.

show abstract

Section: Introductionmentioning

confidence: 99%

A Deep Learning-Based Assisted Teaching System for Oral English

Liu

2022

Security and Communication Networks

View full text Add to dashboard Cite

show abstract

“…Nevertheless, because of the shortcoming that it is very sensitive to its initialization process and training process and easy to overfit (Rothmann and Porrmann, 2022;Matsuo et al, 2022), reinforcement learning should be improved further around the issue of this article. In this respect, the random forest algorithm could be introduced to solve the overfitting problem (Ouadah et al, 2022;Fei et al, 2022).…”

Section: Methods Proposition 41 Algorithm Outlinementioning

confidence: 99%

Section: Methods Propositionmentioning

confidence: 99%

Sharing network features analysis and dispatching strategy design

Shi

2022

IMDS

View full text Add to dashboard Cite

PurposeA vital job for one sharing business is dynamically dispatching shared items to balance the demand-supply of different sharing points in one sharing network. In order to construct a highly efficient dispatch strategy, this paper proposes a new dispatching algorithm based on the findings of sharing network characteristics.Design/methodology/approachTo that end, in this paper, the profit-changing process of single sharing points is modeled and analyzed first. And then, the characteristics of the whole sharing network are investigated. Subsequently, some interesting propositions are obtained, based on which an algorithm (named the Two-step random forest reinforcement learning algorithm) is proposed.FindingsThe authors discover that the sharing points of a common sharing network could be categorized into 6 types according to their profit dynamics; a sharing network that is made up of various combinations of sharing stations would exhibit distinct profit characteristics. Accounting for the characteristics, a specific method for guiding the dynamic dispatch of shared products is developed and validated.Originality/valueBecause the suggested method considers the interaction features between sharing points in a sharing network, its computation speeds and the convergence efficacy to the global optimum scheme are better than similar studies. It suits better to the sharing business requiring a higher time-efficiency.

show abstract

“…Another negative penalty, proportional to the absolute difference between the current and previous actions (as indicated in (7)) is inflicted on the agent to discourage commanding largely different actions, particularly in opposing directions, which results in an oscillatory behavior of the physical platform.…”

Section: E Reward Designmentioning

confidence: 99%

“…Remarkable progress has been registered in developing RL algorithms for various robotic applications including, but not limited to, manipulation [3], navigation [4], [5], and control [6]. Reinforcement learning (RL) is a machine learning paradigm that relies on a rewarding system to train a set of neural networks to make sequential decisions to execute a particular task [7]. More specifically, an RL agent first receives a set of observations parameterizing the state of the environment in which the task is being carried out.…”

Section: Introductionmentioning

confidence: 99%

Learning to Navigate Through Reinforcement Across the Sim2Real Gap

Azzam¹,

Chehadeh²,

Abdulhay³

et al. 2022

Preprint

View full text Add to dashboard Cite

<p>Amid the recent advances in robotics and machine learning, unmanned aerial vehicles (UAVs) have shown evident proliferation across various applications. Consequently, the involvement of UAVs in populated environments has progressively become inevitable, putting forward stringent safety and security measures. In this work, we develop a deep reinforcement learning-based UAV-navigation approach that blends decision making with behavioral intelligence. In particular, a reinforcement learning (RL) agent is trained to instruct the UAV on how to accomplish a goal-oriented task, while assuring the safety of the UAV and its surroundings. Upon arriving at the goal position, the RL agent slows the UAV down preparing it for landing. The safety of the UAV and the environment are attained through a robust collision avoidance capability, embedded into the RL-based navigation system and considers both static and dynamic obstacles in the environment. Training is exclusively carried out in simulation, where a high fidelity UAV controller model is used to perform the simulated maneuver. The proposed approach was tested in simulation and then shown to directly transfer to reality without explicit sim2real gap transfer techniques. Experimental results demonstrated the agent's capability to achieve the navigation task with 90% success rate.</p>

show abstract

Deep learning, reinforcement learning, and world models

Cited by 223 publications

References 51 publications

A Deep Learning-Based Assisted Teaching System for Oral English

A Deep Learning-Based Assisted Teaching System for Oral English

Sharing network features analysis and dispatching strategy design

Learning to Navigate Through Reinforcement Across the Sim2Real Gap

Contact Info

Product

Resources

About