Unfortunately, after the week was mostly spent, I was struggling with some of the technical limitations of the current version of the API, the largest issue being how data is passed to the bot. The project itself had a time duration of two weeks, and I had initially planned on spending one of the two weeks building out a robust action pool for the bot to work off. This is in part, a basis of the action pool the current iteration of my bot uses. Fortunately, there was a major resource in the form of Steven Brown's tutorials in which he took the Python API and built a deep Q-learning agent in a limited action pool.
PROJECT HAVEN SC2 CODE
This led to a less than neat design to the code, and such that it was and still is in very active development, an understanding of the code took some time. This is the C++ API and all others were built on top of this. Due to the nature of the game development, there was an initial API built so that the game could process bot inputs built on Google's Protocool Buffers extensible mechanisms. My only known language is Python so choosing an API was simple. APIs exist for C++, Javascript, and Python.
PROJECT HAVEN SC2 SOFTWARE
The starting point was learning the software stack that I had to use, ie. Since reinforcement learners can learn from their past actions, and the API at this time did not support loading in of replays for imitation learning, despite the fact that large replay datasets did exist, I was unable to make use of those and had to focus on pure reinforcement learning. The starting point of this project is not gathering data, unlike many other data-science projects. The beauty of the reinforcement learner problem is that a good reinforcement learner should be able to learn the game without any programmer domain knowledge and because of that be able to teach itself without any domain specific hacks and also find it's own novel solutions to problems that humans haven't thought of yet. I'm very excited to delve deeper into the world of reinforcement learning, commonly regarded to be the possible holy grail of AI in terms of having the potential to be a general learner, and to possibly make a 'best in the world' player and do the same thing that AI has done to games like Chess and Go. So far, progress by DeepMind has been tight-lipped with few publications apart from the initial statement of using Starcraft II as a testbed for developing AI, but the APIs have been released to the public and projects have been encouraged. In particular, a well-known name in AI and especially reinforcement learning, DeepMind is taking a forefront position by developing APIs to give a computer agent access to play the game. Given the complexity of the game (refer to the bottom section: Introduction To The Game for an overall description of the game), it represents a testbed for testing reinforcement learning artificial intelligence, and given the complexity, it is expected that a successful AI will be revolutionary. Starcraft II Asynchronous Advantage Actor Critic reinforcement learner for PySCII API version 1.1 Updates will come relatively frequently as this is updated. This is a work in progress and the AI can be started by way of calling python deepmind_api_modified/pysc2/bin/agent.py This hasn't been tested on other systems and may not work. Branch has been modified to support the learning Agent. Starcraft II A3C learning agent First off, credits go to Steven Brown, and Jaromir, for their tutorials.Īs is, PySC2 branch version 7a04e74 is installed via anaconda on Ubuntu 16.04.
0 kommentar(er)
