Most consequential problems, whether a game-theoretic policy space, an optimal-control cost surface, a business strategy choice, or a numerical optimization, share the same underlying shape. They are not single-scale terrains. They decompose into nested basins: a few enormous strategic basins, finer tactical valleys conditioned on the strategic choice, and a thin operational roughness on top. The energy-landscape picture is a unifying mental model for this structure. Whatever the domain, the same three levels show up, each with its own barrier heights, its own equilibration timescale, and its own answer to the question “what kind of solution approach actually makes sense here?” The right tool depends on the level: random restarts or simulated annealing for jumping between strategic basins, search or branch-and-bound for navigating tactical sub-basins, and local methods like gradient descent, Newton steps, or line search for the operational scale. The interactive below lets you toggle between the three levels of the same landscape and see what becomes visible at each grain.
- ◈ Strategic. Broad modal contours: choosing which fundamental basin to descend into. Barriers between these basins are enormous. Getting this wrong means no local refinement helps.
- ◇ Tactical. Nested valleys within a strategic basin. The landscape’s shape here depends on which strategic choice was made. Moderate barriers, faster equilibration.
- · Operational. Micro-structure and roughness of the local terrain. Small barriers, rapid convergence. Gradient descent works well here, because the topology is already determined.
The decomposition is
\[E(s, t, o) = E_{\text{strategic}}(s) + E_{\text{tactical}}(t \mid s) + E_{\text{operational}}(o \mid s, t).\]Strategic terms set which basin you live in. Tactical terms shape the sub-basin given that choice. Operational terms are the local roughness. Conditioning matters: the tactical landscape is not a fixed object; it is reshaped by the strategic choice, and the operational landscape is reshaped again by the tactical one.
The same picture across domains
The same three levels recur in very different problem classes. In a game-theoretic setting, the strategic basin is the equilibrium class (cooperate, defect, or some mixed regime), the tactical valley is the policy family within that equilibrium, and the operational scale is parameter tuning of that policy. In optimal control, the strategic basin is the mode (waypoint set, controller family, or maneuver primitive), the tactical valley is the trajectory class within that mode, and the operational scale is the local feedback law. In business strategy, the strategic basin is the market or product category, the tactical valley is the go-to-market or distribution model, and the operational scale is daily execution. In numerical optimization, the strategic basin is the loss-landscape attractor selected by initialization or random restart, the tactical valley is the sub-basin chosen by your inner-loop scheme, and the operational scale is the local gradient step.
The labels differ but the geometry does not.
Why the hierarchy matters
Local search has a different meaning at every level. Gradient descent on the operational scale is fine, even cheap, because barriers are small, equilibration is fast, and the topology around you was already determined by the higher levels. The same gradient descent run at the strategic scale will sit forever in whatever basin it was initialized in, because the inter-basin barriers are huge and the landscape is essentially flat almost everywhere else.
This is the practical content of the “explore vs. exploit” dichotomy. Exploit means descend the operational gradient. Explore means commit budget to changing which strategic or tactical basin you are in, which, on this picture, is barrier-crossing rather than gradient-following.
How it connects to the uncomputable north star
A Bayes-optimal agent in principle treats the whole hypothesis space as one undifferentiated landscape and weighs every possibility. In practice no embodied agent can. Approximation choices get organized along exactly this hierarchy: the prior coarsens at the strategic scale, the inference scheme gets refined at the tactical scale, and the planning budget is mostly burned at the operational scale. The shape of a working agent is the shape of how it allocates compute across these levels.