Group items tagged

Evolutionary Strategies are able to explore broader areas of the search space than reinforcement learning techniques. Thus, they are able to encounter strange bugs resulting in large rewards.

It will be the new hype in a few years when DL is settled....

Prediction-based method for encouraging reinforcement learning agents to explore their environments through curiosity (reward for unfamiliar states). Learns some games without any extrinsic reward!

Fun failure case: agent gets stuck in front of TV.

Not read this article but on a related note: Curiosity and various metrics for it have been explored for some time in robotics (outside of RL) as a framework for exploring (partially) unfamiliar environments. I came across some papers on this topic applied to UAVs when prep'ing for a PhD app. This one (http://www.cim.mcgill.ca/~yogesh/publications/crv2014.pdf) comes to mind - which used a topic modelling approach.

Correct me, if I am wrong, but AFAIK this is the first time an AI enters a ladder, i.e. playing against humans on their own terms in the wild and not as part of some pre-arranged experiment.

Another victory for our future robo-overlords.

:) always thought so!

The direct link without the Washington post paywall https://www.researchgate.net/publication/329131866_Going_With_Your_Gut_The_Inaccuracy_of_Forecast_Revisions_in_a_Football_Score_Prediction_Game

Actually read the paper now and think that this is very doubtful that this could be generalised. They used predictions for football games by those betting on outcomes ...