akashin / gail Goto Github PK

I propose to use Google Test (https://github.com/google/googletest), but if there are any better proposals, I'm happy to listen.

Надо загрузить существующие реализации на C++ для того чтобы генерировать обучающие данные

There is a cool derivative-free method based on evolution algorithms, that can be applied to optimize the parameters of the player policy. It would be cool to try out this method on the CodeVsZombies task and see if it works. Then we can separate it and make it applicable in other tasks.

Here is a good lecture about this methods: https://www.youtube.com/watch?v=SQtOI9jsrJ0&feature=youtu.be
And an article with references https://blog.openai.com/evolution-strategies/

We want to have some good solution that can be used as an opponent or generate training data for the RL-based agent.

One of the methods that seems to be applicable in CodeVsZombies game and in many strategy games is hierarchical learning. The general idea being that we represent the agent as a high-level controller that operates by every step picking a set of sub-controllers that are relevant to the current situation, and the sub-controllers solve one particular task (destroy a cluster of zombies, protect human) well. Both controller and sub-controllers can be trained from data, but the hybrid scheme in which sub-controllers are implemented by hand, and high-level controller is trained are also possible.

It would be great to try such method for CodeVsZombies task.

Here is the article that gives a demo of the technique: https://blog.openai.com/learning-a-hierarchy/

Also, there are quite a few references here: http://realai.org/hierarchical-learning/
In particular FeUdal Networks seems to be a working approach.

At the moment all players are implemented in C++.
However, nothing prevents us from implementing an ExternalProcessPlayer/PythonPlayer that will under the hood interact with separately running process that will make decisions about what actions to take. This will allow us to implement players in any language.

When designing this, one should have in mind the cost of interaction between processes (File? Pipe? Shared Memory?) and the ability to extend this approach to other languages (use some standard message protocol, e.g. json, msgpack or protobuf).

We want to have simulator for this game: https://www.codingame.com/multiplayer/optimization/code-vs-zombies

We already have a client and a player implementation.
Interesting properties of the game:

• Big discrete action space
• 1 player
• Set of fixed evaluation levels

It would be awesome to have a neural network training in the toolkit, to be able to apply it in different contexts. One of the first demos might be playing the FantasticFour game with neural network that picks the next move given the state of the board.

This can be achieved through the following steps:

• Find or write a C/C++ neural network library (see https://github.com/attractivechaos/kann)
• Implement a network that takes board as an input and returns a next move
• Train the network to predict the moves of one of the existing algorithms (e.g. simple TreeSearch)
• Train the network through reinforcement learning to play on it's own (you might find this article useful http://karpathy.github.io/2016/05/31/rl/)

Bonus: Check out AlphaGo Zero paper (http://tim.hibal.org/blog/alpha-zero-how-and-why-it-works/) and implement it.

Monte-Carlo tree search is a great tool that is used in many successful game-playing agents (e.g. AlphaGo). It would be useful to have it in a codebase in a way that will be adaptable for different games. It may also be integrated with reinforcement learning in future.
Here are relevant posts about the topic:

• http://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/
• https://blog.theofekfoundation.org/artificial-intelligence/2016/06/27/what-is-the-monte-carlo-tree-search/

There is a nice paper that should serve as a reference to different applications of MCTS:
http://mcts.ai/pubs/mcts-survey-master.pdf

Also paper explaining how it's used in AlphaGo Zero might be useful: https://github.com/B-C-WANG/AlphaGo-Zero-Paper

Я предлагаю взять вот эту игру, т.к. по-моему мы оба её решали, и в ней довольно простое пространство действий:
https://www.hackerearth.com/problem/multiplayer/fantastic-four/description/

Надо будет запрограммировать окружение для игры чтобы можно было проводить симуляцию.

Add test witch test speedups deterministic algorithms is correctly

• TreeSearch
• TreeSearch with caching
• AlphaBetta
• AlphaBetta with caching
• AlphaBetta with move order heuristic