Mixture of Experts
Many specialists plus a router, instead of one generalist.
A single policy trained across every market condition has to average over conflicting behaviours. If a calm market rewards patience and a volatile market rewards speed, one network trained on both learns something mediocre at each.
The Mixture of Experts idea, going back to Jacobs, Jordan, Nowlan & Hinton (1991), is to stop asking one model to do everything. Keep $K$ specialist policies, and add a router that decides which specialist handles the situation in front of you.
Hard versus soft routing
Hard gating picks exactly one expert — winner takes all. Clean, interpretable, and each expert trains only on its own slice of the data.
Soft gating blends every expert's output, weighted by relevance. Smoother, differentiable end-to-end, and no expert ever starves for data.
The trade-off is real: hard routing gives you specialists but fragments your training data $K$ ways. With 2,500 training episodes and $K = 8$ experts, each expert sees only a few hundred. That data fragmentation is a genuine cost, and it's one of the things my research tries to measure rather than hand-wave.
Routing without looking ahead
If the router picks an expert using information from the episode it's about to trade, you've leaked the future into the past and your backtest is fiction. The safe construction routes on the previous episode's features — you decide which specialist to use based on what the market just did, never on what it's about to do.
Modern large language models use the same architecture at enormous scale (Shazeer et al., 2017), for a different reason: sparsity. If only a couple of experts activate per token, you can grow total parameter count without growing the cost of a forward pass.