5

We know what Lee's strategy was during the game, and it seems like the sort of thing that should work. Here's an article explaining it. Short version: yes, we know what went wrong, but probably not how to fix it yet. Basically, AlphaGo is good at making lots of small decisions well, and managing risk and uncertainty better than humans can. One of the things ...


4

Dueling architectures create bigger differences in the values of actions in the state space. This is because the state-value V(s) function is estimated separately from the state-action value Q(s, a). A new quantity, the advantage of an action, can then be defined as A(s, a) = Q(s, a) - V(s). The Q function, however, measures the the value of choosing a ...


4

In SC2, players have more control over every minuet mechanic (constructing buildings, resource mining and management, controlling minions...) in the game, thus putting more tactical responsibility on the burden of the player. In DOTA2, the player is only in control of the super-powered hero itself, and not much of the other aspects of the gameplay. It is ...


4

Those AI-learning programs may have very similar scheme. We are changing only inputs and possible actions (like "use skill" or "move here"). Starcraft AI must do a lot of actions and control many units. Dota is MOBA so bot should be good in positioning on map for example. Different opponents to destroy and target for win. AI needs to play many games for ...


3

Based on this comment in the Issue I created about this question on github, it looks like there is confirmation that at least DeepMind does not use this kind of functionality in their Atari experiments, contrary to what is implied by the comments in the OpenAI baselines code.


3

It means that there is no explicit coding of action choices to promote to queen, it is the default assumption if the underpromotion actions are not taken. The Alpha Zero chess implementation can represent promotion to queen by not selecting an underpromotion action, whilst moving a pawn so that it qualifies for promotion.


3

In their blog post, they link to (among many other papers) their IMPALA paper. Now, the blog post only links to that paper with text implying that they're using the "off-policy actor-critic reinforcement learning" described in that paper, but one of the major points of the IMPALA paper is actually an efficient, large-scale, distributed RL setup. So, until ...


3

Although what @Jaden said may be true by itself, it does not really serve to answer my question as I have seen after conducting numerous experiments, and finally reaching close to Dueling Network performance using a normal Double DQN (DDQN). I made the following changes to my code after closely examining the OpenAI baselines code: Used PongFrameskip-v4 ...


2

Switching required a re-learn. Also, note that: We use the same network architecture, learning algorithm and hyperparameters settings across all seven games, showing that our approach is robust enough to work on a variety of games without incorporating game-specific information. While we evaluated our agents on the real and unmodified games, we ...


2

For example DeepMind [Lab] is not a standard or Open Source friendly I'm not sure where you got that info from... as far as I'm aware, DeepMind Lab is definitely used in various publications (maybe primarily publications from DeepMind, but still). Considering the github repo has the GNU GPL 2 license, it also seems Open Source-friendly to me. Another ...


1

The formula in question uses a function N(state, action) that defines a visit count of a state-action pair (introduced on page 3). To describe how it is used, lets first describe the steps of AlphaGo Zero as a whole. There are 4 "phases" to the Monte-Carlo tree search in AlphaGo Zero as depicted in Figure 2. The first 3 expand and update the tree and ...


1

DeepMind used MuJoCo (see also the related paper MuJoCo: A physics engine for model-based control) for the simulations, as they stated in section 3.1 of their paper Emergence of Locomotion Behaviours in Rich Environments (2017), which is the paper you should read to know more about their results related to those animations of skeletons that try to walk or ...


1

The thing is, the decoder samples from a latent mu and sigma, so you cant sample from a raw image directly. But if you’re trying to put a random image into the encoder of a trained VAE to match it to some sample image (via reconstruction loss), then your random input image will converge to the target sample. This will work when the following VAE ...


1

Basically, it means that the "localization network" should output a set of real valued parameters (typically 6 numbers). The word "regression" doesn't bear any specific meaning. Any network that relies on the original image as input (directly or indirectly) and outputs 6 numbers, would work. And its last layer would qualify as "regression layer" as long as ...


1

Switching requires relearning, the network did not have a single set of weights that allowed it to play all games well. This is due to the catastrophic forgetting problem. However, recent work has been done to overcome this problem: "Overcoming catastrophic forgetting in neural networks", 2016 Paper: https://arxiv.org/pdf/1612.00796v1.pdf


Only top voted, non community-wiki answers of a minimum length are eligible