Create MCTS trained agent

[hmosousa]

First implementation:

• keep the reward ratio (reward / max reward) for each playout instead of the number of wins
• use a random agent in the sampling process (light sampling)