tag:blogger.com,1999:blog-9275314.post3210338334882596740..comments2024-03-23T11:34:23.998-04:00Comments on Togelius: Which games are useful for testing artificial general intelligence?Julian Togeliushttp://www.blogger.com/profile/09333191187316058782noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-9275314.post-65864290401326960322016-12-06T10:51:05.081-05:002016-12-06T10:51:05.081-05:00Ah, if only there was big money to be made in this...Ah, if only there was big money to be made in this area today, I'd drop my job... You're so on the right path, if we live full Ai, you'll be at the fore front.ciro.santillihttps://www.blogger.com/profile/02642132378416878124noreply@blogger.comtag:blogger.com,1999:blog-9275314.post-92208819268432851482016-07-18T09:23:32.036-04:002016-07-18T09:23:32.036-04:00I feel you missed some important benefits of the V...I feel you missed some important benefits of the VGDL framework over the ALE:<br /><br />Firstly, with VGDL you get given the objects themselves. This is great for developing techniques likely to be useful for the games industry, as very few developers are going to care about truly end-to-end systems from visual input. This may also be important to research as well, allowing us to bypass having to deal with low-level, noisy information for certain experiments. I know of at least one paper where an object-grid input was used as features for a learner, something trivially extractable when working within the VGDL framework.<br /><br />Secondly, with VGDL you have access to the ground truth of the model of the game. This is potentially useful for validating any model-learning algorithms. Alternatively, it can be used to skip the model learning step to test other algorithms which may rely upon a full model (as opposed to a one-step model).<br /><br />Finally, and you touch upon this briefly, VGDL makes it possible to synthesise your own games. While this is not necessarily the best of ideas from an experimental point of view, it makes it possible to generate datasets which might not be available otherwise. Really want to test your general platformer algorithm, but current titles are too difficult to interface with? Replicate them in VGDL!<br /><br />I think the true potential of the VGDL as a research framework has not yet been tapped, particularly for learning. The ALE is a very nice, and rigorous, way of testing and demonstrating advanced AI techniques, but it's not always the most practical platform to test on. As well as being a great platform for developing AI with practical applications, some of the short-cuts VGDL provides are a great way of making investigating problems of the future easier. I'd love to see it integrated as an experimentation platform into something like the OpenAI Gym (gym.openai.com) in future.Anonymoushttps://www.blogger.com/profile/12324981280936660390noreply@blogger.com