Warwick Annual Retreat Projects Pommerman – Multi-Agent Learning Competition

3 posts 2 followers RSS feed

1. Henry Charlesworth

   #1 21:54, Fri 23 Feb 2018

   Pommerman – Multi-Agent Learning Competition

   Organised by -
   Henry Charlesworth

   
   Abstract -
   Pommerman is an implementation of the SNES game “Super Bomberman”designed such that it can be used as a playground for multi-agent learning. This is a very simple four player game with two different game modes (FFA and Team 2v2) but is made complicated by the fact that it is adversarial and you are competing against other agents. The idea of the Pommerman project is that people from around the world will try and train agents to play this game, submit their entries and then the organizers will host competitions where these all compete against each other.

   
   Aims and Objectives -
   Realistically it will be difficult to create a highly competitive agent within the amount of time we have but hopefully it will be possible implement something simple that can at least play the game sensibly and then to brainstorm ideas that we could try to implement afterwards if people are still interested.

   
   Of Interest to -
   People interested in machine learning (particularly reinforcement learning) and people who like old SNES games. I am mainly interested in doing this because I’d like a chance to try and implement some reinforcement learning algorithms.

   
   Resources Necessary -
   Agents can be developed using any framework essentially (the only requirement is that you submit a Docker container). I was planning to use tensorflow but am open to suggestions from people who know a lot more about machine learning than I do.

   
   References -
   https://github.com/MultiAgentLearning/playground
   for the very keen: https://github.com/LantaoYu/MARL-Papers

    

   2 likes
    
2. Edoardo Barp

   #2 14:19, Sun 4 Mar 2018

   This is definitely somethign  I'm interested it! 

   I've never implemented or tested any reinforcement learning, although I now have some understanding (thanks to a module) of basic reinforcement learning methods (Q-learning, MC tree search..).

   How were you thinking of using tensorflow? I've only used it for neural networks, which I thought was the only thing it did. Were you planning on training a NN using episode learning for instance?

   Great project anyway!

   0 likes
    
3. Henry Charlesworth

   #3 11:59, Mon 5 Mar 2018

   OK cool, glad to hear it! I'm also not at all an expert in reinforcement learning but it's an area that I've learning about recently and getting pretty interested in. I've had a go at implementing some Deep Q networks and simple policy gradient algorithms (but my understanding is that these don't actually work all that well in multi-agent environments) so we probably need to do something a bit different. Still I was thinking tensorflow because probably training a neural network of some kind is the only approach I can think of at the moment (e.g. maybe trying to learn some more about the competitive self-play methods they've been using at OpenAI). I'm definitely open to suggestions though!

   0 likes