- Home /
training ml agent in alkkagi game
Hi, I've made an environment using ml-agent package, for a 2-player game called alkkagi, which is basically a game about shooting the pieces on my side and making the opponent's piece to fall off the plane. If every piece on one side falls off, the game ends.
The problem is that the agents aren't learning anything, even after training it a lot. I really need help on this.
The agent has
one discrete action, which decides which piece to shoot, and
two continuous actions, which each decides the power and direction of shooting the piece chosen by discrete action above.
I'm not sure that this method about using discrete and continuous action is correct.
I've simply gave every pieces' position and boolean value(dead or not) as the observations. I guess this provides enough information, but I'm not sure. Will it be the observation's problem?
The reward is given as +1, 0, -1 when it wins, ties, and loses the game.
I will upload my configuration in png. (txt is not showing)
I trained the agent by using self-play option.
Your answer
Follow this Question
Related Questions
Itween path constraint with multiple vehicles at once 1 Answer
Draw path along Navmesh agent path 3 Answers
Nav Agent multiple Units problem 0 Answers
Access to scriptable objects 1 Answer
Unity stutters when training with Machine Learning Agents. 0 Answers