A {\it pattern} is a string of constants and variables, 
and is {\it regular} if each variable occurs in the 
pattern at most once. 
The present paper deals with the problem of learning 
languages generated by regular patterns with two kinds 
of variables, {\it erasing} and {\it nonerasing} from 
positive examples within Gold's model.   
First, we  show an equivalence theorem of a semantic 
containment $L(p) \subseteq L(q)$ and a syntactic 
containment $p \preceq q$ for regular patterns $p$ and $q$. 
Then we show that the class of regular pattern languages 
is polynomial time inferable from positive examples.