add alternating blue/red learning phases

[martyn-smith]

We might see significantly higher learning stability if we force blue to have a learning phase against a static red policy, and vice versa. Easy enough to implement, possibly a new branch.