Fixes an error related to depth-limit in grammars without terminal rules

[ancorso]

When we have a grammar where not every type has a terminal rule, setting a depth limit when sampling expressions can cause an error. For example, the grammar

grammar = @grammar begin
        A = B
        B = C
        C = D
        D = E
        E = 1
    end

will cause an error when we sample with a depth limit of 1 i.e. rulenode = rand(RuleNode, grammar, :A, 1).

I changed the logic so that if the depth limit is reached and a terminal rule is available then choose that rule. If no terminal rules are available then it chooses from the available rules. The example above is added as a test.

[coveralls]

Coverage decreased (-1.3%) to 86.667% when pulling 3002dab on ancorso:depth_lim_iss into 8573b55 on sisl:master.

[mykelk]

Looks good to me!

[rcnlee]

Thanks Anthony! So instead of throwing an error, it'll return the incomplete expression and the user should handle it on their side (e.g., in the reward function)? How will the user know that the expression is incomplete and need special handling?

The other question I had was regarding performance, since rand is a core inner loop function. The filtering is now happening at every step (rand is recursive) even when majority of the time maxdepth > 1 and it is not needed. Perhaps we can restructure the if-statement to avoid that.

[ancorso]

Hi Ritchie,

It does complete the expression, the expression may just have a depth larger than max_depth. The goal is to always output a valid expression.

As for performance, I can rework the if statement so that the filtering only happens when we reach the max depth. It'll just be an extra line or two.

[rcnlee]

Sounds good, I agree that is the best way to handle it. I'll merge when the checks complete. Thanks Anthony!