Multiple categorical IVs and DVs with 3/5 levels prediction model I have a data-set with 
8 categorical IVs with 2/3 levels (one level for one type of conditions), 
2 categorical DVs with 3/5 levels (one level for one type of responses:dis/even/ad). 
Participants were tested repeatedly. 
I am looking for a best model in R for prediction. 
So far, the most relative approach I've found online is multinomial-logistic-multilevel-models-in-r, and I used 'brms' package: 

############### model for all 6 effects

brm2.1 <-brm(allocation~age_group+gender+country+competition+relation+cost)+(1|id),data=R2long,family=categorical)
I managed to get some outputs like the following: 

############### summary(brm2.1)



############### I also ploted the model:

plot(brm2.1,N = 5)

Multinomial-logistic-multilevel-models is new to me so I don't know how to interpret the summary outputs, i'd like to know information like:


*

*1: how to find the best fitted model? 
I did modelling for 'glmer' and i wonder if that's the same way, like top down bottom?

*2: how to explain the effects of each IV and interactions? like the effect size, significance,  affect the whole model, etc.

*3: how to find the predicted percentages? like how many people may respond in 1 of the 3 ways to certain manipulation? 

*4: what does the plot mean?? i never see plot like this before.....
I know that're a lot questions, i am really grateful any help. 
 A: As mentioned in comments, you used a Bayesian package, perhaps without specifically knowing that it was a Bayesian package. I'm not very familiar with Bayesian analysis, and with no offense intended, it appears you are not either. So it would probably be better to switch to a different package. I'm not familiar with R packages, but the UCLA stats webpage offers a worked example of multinomial regression with the nnet package.
In your question, you fit a model for the multinomial regression to the dependent variable allocation, which takes the levels dis, even, and ad. Here, the package chose ad as the base level. You got two sets of coefficients representing the effect of a unit change in each variable on the log odds (or something) of choosing that category versus the base category. I said "or something" because you appear to have applied an identity link, which is not the default in real multinomial regression (but I guess you can do this in a GLM). Identity link implies that your coefficients are on the probability scale, but that doesn't really look possible given some of the coefficients (e.g. some are over 1), so I have no idea what scale the output is on. So much for answering your question 2. Whatever the package you use, your output is going to come in a similar format because that's what multinomial regression does.
Question 4: those are diagnostic plots. Real Bayesians, please forgive me if I mangle this, but my understanding is that in Bayesian estimation, you are essentially simulating what the parameter values of interest could be. The plots on the right are a trace of the parameter values in each simulation. You basically want the estimation algorithm to wander randomly over the parameter space close to whatever the Bayesian equivalent of the MLE estimate is. It appears that those 5 plots meet that criterion, but you only asked brms to plot 5 parameters. I believe the plots on the left are the posterior distribution of the parameter. You want these to look approximately normal. The ones you presented do. These are just model diagnostics you should perform. If you have no idea what any of this means, then (as mentioned earlier) switch packages. That sort of takes care of your Q4.
For your Q1, you only fit one model, so there's no best fit model. In Bayesian estimation, there are parallels to the information criteria we typically use in frequentist maximum likelihood estimation.
I've never used that package, so I haven't even attempted to answer Q3.
