Implement heuristics for restricting design space

[drbenvincent]

Following on from #2, I need to implement the heuristic component for DARC tasks. This takes the set of all possible designs and shrinks the set of designs to be entered into the design optimisation procedure on a trial-to-trial basis.

The plan is still to implement what Tom wrote here in Matlab. However, that was based on all models being subjective value models, and so you could compute VA and VB for each prospect. Tom's code assumes we can access the design variable VB-VA and also VB+VA. But this is no longer the case when we consider our broader class of models... we now have attribute-wise models which simply don't operate by assessing the subjective value of each prospect. But all models do have a decision_variable axis... just not necessarily an easy way to specify orthogonality from the decision_variable.

• [x] Removal of designs which we've already run. This is optional because the experimenter might specifically want to allow repeated designs.
• [x] Removal of designs with highly predictable outcomes. Why? Because we do not expect to learn much for these trials. If a participant did respond oppositely from what we'd predict, this is more likely to be a lapse trial than a genuine learning opportunity.
• [ ] Check/think about ability to get pChooseB absent from effects of epsilon

Stuff to be addressed in other Issues

• Implement the heuristic_order approach Not sure if this was deprecated?
• Implement the subjective_value_spreading code. I believe this was the latest, best approach?

[drbenvincent]

The heuristics I've added so far (removal of already asked designs, removal of highly predictive designs) has already helped a lot. Doing a few tests shows a big improvement in the design selection and tightness of the posterior

[drbenvincent]

Basic set of heuristics done. Need to implement the more clever stuff. See #27