# Results of grid-search for $d=3$

The best set of parameters seem to be pop_size=100, n_gen=100, ngames=100, eps=0.1


Note that MWPM is beaten consistently for some specific set of parameters!

The curves roughly stay with the same ordering against the error rate, this motivates the visualization at a specific error rate (taken to be 0.1 in the following).

## Influence of the population size

**Comments**
- Increasing the population size is always beneficial (except this one case with low epsilon)

## Influence of the number of generations

**Comments**
- In most cases, increasing the number of generations from 100 to 200 increase marginally the efficiency of the correction
- For two cases, it seems to make things worse
- **100 generations seem sufficient to get a solution**

## Influence of the number of training puzzles

**Comments:**
- There is no clear trend, analyze better?

## Influence of epsilon $\epsilon$

**Comments**
- not so clear

# Results of grid-search for $d=5$