Proximal Policy Optimization (PPO) is widely regarded as one of the most successful deep reinforcement learning algorithms, known for its robustness and effectiveness across a range of problems. The PPO objective encourages the importance ratio between the current and behavior policies to move to the "right" direction -- starting from importance sampling ratios equal to 1, increasing the ratios for actions with positive advantages and decreasing those with negative advantages. A clipping function is introduced to prevent over-optimization when updating the importance ratio in these "right" direction regions. Many PPO variants have been proposed to extend its success, most of which modify the objective's behavior by altering the clipping in the "right" direction regions. However, due to randomness in the rollouts and stochasticity of the policy optimization, we observe that the ratios frequently move to the "wrong" direction during the PPO optimization. This is a key factor hindering the improvement of PPO, but it has been largely overlooked. To address this, we propose the Directional-Clamp PPO algorithm (DClamp-PPO), which further penalizes the actions going to the strict "wrong" direction regions, where the advantage is positive (negative) and importance ratio falls below (above) $1 - β$ ($1+β$), for a tunable parameter $β\in (0, 1)$. The penalty is by enforcing a steeper loss slope, i.e., a clamp, in those regions. We demonstrate that DClamp-PPO consistently outperforms PPO, as well as its variants, by focusing on modifying the objective's behavior in the "right" direction, across various MuJoCo environments, using different random seeds. The proposed method is shown, both theoretically and empirically, to better avoid "wrong" direction updates while keeping the importance ratio closer to 1.

By virtually every key metric, efforts to fight climate change are going too slowly, according to findings by a coalition of climate groups. In some cases, things are moving in the wrong direction. An eroded iceberg is seen is seen floating near Horseshoe Island, Antarctica. In the 10 years since the signing of the Paris Agreement, the backbone of international climate action, humanity has made impressive progress. Renewable energy is increasingly cheap and reliable, while electric vehicles are becoming better every year.

Originality: While lots of works have studied the property of the endpoint found by SGDs, the literature looking at the SGD training dynamics in the context of deep neural networks is sparser, and the loss contribution metric appears novel to me. The paper is therefore original from that aspect. Quality: The paper is in general of good quality. However, few specific points could be improved: - It would be nice to characterize the approximation errors introduced by the first order taylor expension - Authors claim that the Loss contribution is grounded while other Fisher information-based metrics heavily depends on the parametrization chosen. Could the authors expend on this point and provided a more detailed comparison between LC and the metrics introduced in [1] and [13] - In the introduction, authors claim that entire layers drift on the wrong direction during training.

While we have all been focused on facial recognition as the poster child for AI ethics, another concerning form of AI has quietly emerged and rapidly advanced during COVID-19: AI-enabled employee surveillance at home. Though we are justifiably worried about being watched while out in public, we are now increasingly being observed in our homes. Surveillance of employees is hardly new. This started in earnest with "scientific management" of workers led by Frederick Taylor near the beginning of the 20th century, with "time and motion" studies to determine the optimal way to perform a job. Through this, business management focused on maximizing control over how people performed work.

A driverless train that injured 14 people in Yokohama on Saturday after moving in the wrong direction continued moving for 1 meter even after hitting a buffer at a station because of the way the buffer works, the train operator said Monday. Saturday's accident occurred at Shin-Sugita station on the Kanazawa Seaside Line. Of the 14 passengers hurt, six sustained serious injuries. According to the operator, Yokohama Seaside Line Co., the unmanned train traveled for 25 meters in the wrong direction, hit the buffer, which is designed to absorb any impact, and then continued to move for about a meter. The Japan Transport Safety Board and the operating company are specifically investigating the circumstances of the accident, and believe that the impact was magnified when the moving train hit the buffer.

YOKOHAMA - A probe was launched Sunday into an accident involving an automated train the previous night in Yokohama which left 14 hurt, including six with serious injuries. Officials from the Japan Transport Safety Board began examining the train cars at Shin-Sugita Station on the Kanazawa Seaside Line. Investigators are considering a charge of professional negligence resulting in injuries, police sources said. At 8:15 p.m. on Saturday, an automated train operated by Yokohama Seaside Line Co. traveled in the wrong direction for about 25 meters at a speed of 6 kph. The train had been set to depart the station but instead moved backward and crashed into a buffer at Shin-Sugita Station.

YOKOHAMA - An automated train operated by Yokohama Seaside Line Co. on Saturday traveled in the wrong direction, causing about 20 people to be injured, a local fire department said. Some appeared to have suffered serious but non-life-threatening injuries as the train made contact with a buffer stop at Shin-Sugita Station, the department said, but other details were not immediately available. The trains are on an automated guideway transit system connecting Shin-Sugita and Kanazawa Hakkei in Yokohama.

Artificial intelligence sounds cool in theory, and as Google CEO Sundar Pichai said at the Google I/O event on Wednesday, the company wants to "help you get things done" with AI. But one example that Google used to showcase its AI at Google I/O on Wednesday was anything but exciting. Using its new messaging app called Allo, Google showed how easy it is to find restaurants and make reservations, or find a movie and buy tickets. Allo is designed so you can do those things by having a conversational texting session, as you would with a friend, with a bot called @google that uses the company's new AI platform called Google Assistant. There were some other examples, like recognizing the context of messages and pictures you send between your friends and coming up appropriate short replies so you don't have to come up with the response yourself.

Artificial intelligence sounds cool in theory, and as Google CEO Sundar Pichai said at the Google I/O event on Wednesday, the company wants to "help you get things done" with AI. But one example that Google used to showcase its AI at Google I/O on Wednesday was anything but exciting. Using its new messaging app called Allo, Google showed how easy it is to find restaurants and make reservations, or find a movie and buy tickets. Allo is designed so you can do those things by having a conversational texting session, as you would with a friend, with a bot called @google that uses the company's new AI platform called Google Assistant. There were some other examples, like recognizing the context of messages and pictures you send between your friends and coming up appropriate short replies so you don't have to come up with the response yourself.

Artificial intelligence sounds cool in theory, and as Google CEO Sundar Pichai said at the Google I/O event on Wednesday, the company wants to „help you get things done" with AI.