# Conclusion
## Review
In this project, the main goal is imitating a human mind, and chess is our fisrt experimental field. The imaitation step is split into 2 step, imitating the choice of picking and imitating the choice of moving. Using these two prediction information, the combination combined them and outputs the final prediction move.
## Methods
For each steps, these method had been tested. The chosen method is tagged as :+1: .
* Input Format
* 2D
* 3D :+1:
* Time sequence
* Blending Information (Moving Model's input)
* Hidden state blending (Moving Model's input)
* Picking Model
* Sperate prediction on row and column
* Conv3D mapping :+1:
* LSTM
* Resnet
* Moving Model
* Decoder base
* Conv3D Resnet
* Conv3D + Decoder Resnet :+1:
* Combination
* Pick prediction value first
* Stockfish reprocess :+1:
* Loss approach
## Performance
Here is the best performance among all the player :


## Comparison
Compared with the performance between 1 step imitaion ai **Maia**,the 2 step imitation is defeated by about 23% accuracy. It is quite frustrated to me. There may be some reason cause our Ditto being defeated. The main reason should be the limitation of relatively small database, since ML is a data-driven technique. The model architecture may also be the reason, because this model has't been evaluated in a technically way and even not in shape of classical model.
## Application
Due to the unexpected low performance, there is still a long path and mutiple task waiting for us. Imitating a specific chess player is currently impossible, since a perfect imitaion requires a high matching accuracy. If the accuracy isn't high enough, the error will stack and grow through the chain behavior.

However this project can also be used in training process. Since the top 5 choice pool has a acceptable accuracy, one can take this prediction as a assistant tool and form the final prediction by human brain.

This the very basic chess game developed for showing and analyzing the performance of the model. It can also be used as a training tool.
## Conclusion
In this project, the main discovery should be the usage of Conv3D which weights not only the position but also piece type. An adaptive weighting for feature is important on imitation task. During imatation task, splitting the task into mutiple small task might be helpful since a classical method could be applied between the tasks.

In the future, the combination method may be a citical task. Designing a loss that assign some feature on the answer may be helpful. In the other field, one can also design a model based on our discovery.