I think this is a really good idea. Where TCP will often just multiplicatively decrease sending rate at every dropped packet (which happens 0.5% to 1% of the time if the network is normal), PCC does not react immediately but rather feeds this into a learning algorithm, which also learns from doing frequent experiments increasing/decreasing the sending rate and measuring the impact on loss rate, throughput and latency.

Essentially, if I understand it correctly, it's just measuring loss rate, throughput and latency and iteratively adjusting sending rate in response. I think it's very similar to LEDBAT and delay-based congestion control except that it's not just taking delay from bufferbloat into account, but also loss rate and throughput, which makes sense if the goal is to optimize these.