Adaptive Landscapes

Planned units

• 03
  

  Mount Fuji Landscapes: Independent Parts Make One Smooth Climb

  The simplest landscape assumption: each component contributes to the total score independently, with no interactions between parts. In the additive, tie-free case, there are no suboptimal local peaks, so every improving step leads toward the single global optimum. This baseline is the starting point against which every later model measures what it adds.

  Planned / 60 min
• 04
  

  Rough Mount Fuji Landscapes: A Smooth Trend Plus Local Roughness

  Adds random noise to Mount Fuji's smooth additive trend. A broad improvement direction still exists, but local bumps make nearby scores imperfect guides. Many measured molecular fitness landscapes appear intermediate in this way: a broad additive trend plus epistatic deviations, sitting between pure smoothness and pure randomness.

  Planned / 60 min
• 05
  

  Rugged Landscapes: Interdependent Parts Can Create Local Traps

  Makes components interdependent: changing one part changes the value of others. When these interactions are strong enough to produce sign epistasis (where a change helps on one background and hurts on another), hill-climbing can get stuck on local peaks. This captures interdependence in biology, complementarities in organizations, and architectural coupling in technology.

  Planned / 60 min
• 06
  

  N-K Landscapes: K Controls Coupling Between Parts

  Gives interdependence a tunable dial: N components with K dependencies each. In the standard random NK model, increasing K tunes expected ruggedness from additive, Fuji-like landscapes toward House-of-Cards-like uncorrelated landscapes. This lets you match the model to a specific system rather than treating ruggedness as all-or-nothing.

  Planned / 60 min
• 07
  

  House-of-Cards Landscapes: Neighboring Scores Are Uncorrelated

  The extreme limit of ruggedness: neighboring configurations have uncorrelated scores. Hill-climbing still finds a local peak, but each local comparison gives almost no information about what lies beyond the immediate neighborhood. Search becomes purely myopic, offering little assurance of reaching a globally good solution, so strategies like random restarts, parallel exploration, or recombination become essential.

  Planned / 60 min
• 08
  

  Needle-in-a-Haystack Landscapes: Success Is Rare, Not Misleading

  Introduces a different axis of difficulty from ruggedness: solutions exist but are vanishingly rare in a vast flat space. The challenge is sparsity rather than misleading peaks, and brute force fails because each added dimension multiplies the search space without adding gradient information.

  Planned / 60 min
• 09
  

  Neutral Plateau Landscapes: Equal Scores Still Move Search

  Shows that flat regions of equal value are not dead ends. Sideways movement across a plateau can reposition a searcher so that new uphill moves become reachable, provided the plateau connects to those moves. Neutral drift in evolution and organizational slack in innovation illustrate how lateral steps open doors that pure optimization cannot.

  Planned / 60 min
• 10
  

  Neutral Network Landscapes: Many Equal-Function States Stay Connected

  Extends neutral plateaus by connecting equal-value states into traversable networks. A system stays functionally stable while exploring many underlying configurations, creating robustness because many changes preserve function and evolvability because different network positions border different adaptive possibilities.

  Planned / 60 min
• 11
  

  Holey Landscapes: Viability Replaces a Single Fitness Summit

  Replaces the idea of a single fitness summit with viable versus inviable regions. In high dimensions, viable states can form large connected networks around inviable holes without ever crossing a valley, provided the viable set is sufficiently connected. The question shifts from "how do I reach the peak?" to "how do I stay within the connected viable region?"

  Planned / 60 min
• 12
  

  Fitness Seascapes: Fitness Landscapes Move Over Time

  Drops the assumption that the terrain holds still. The value function shifts as environments, therapies, or competitors change, so the best solution today may be harmful tomorrow. A "dancing landscape" in Scott E. Page's sense, where search must track a moving target rather than converge on a fixed summit.

  Planned / 60 min
• 13
  

  Coevolutionary Landscapes: Other Agents Move Your Terrain

  Where seascapes model exogenous landscape change (environments shift from outside), coevolutionary landscapes make the change endogenous: other agents' adaptations reshape your terrain, so optimization pursues a target that moves because you moved. Another "dancing landscape," driven from within by the Red Queen dynamic of mutual adaptation.

  Planned / 60 min
• 14
  

  Energy Landscapes: Lower Energy Means Greater Stability

  Maps the landscape framework to physics by inverting the value direction: lower energy means greater stability. In protein folding, the canonical example, this means lower free energy under specified conditions. A rugged funnel biases the chain toward its native structure, but kinetic barriers determine whether it can get there, so accessibility matters as much as optimality.

  Planned / 60 min
• 15
  

  Attractor Landscapes: Basins Pull Systems Toward Stable States

  Maps the framework to dynamical systems: basins replace peaks, ridges become tipping thresholds, and the question shifts from "where is the summit?" to "which basin are we in, how deep is it, and what shock would tip us into a different regime?" Sometimes the landscape is a true potential function; sometimes it is a visualization of basin structure. Lakes, ecosystems, and social systems all exhibit this basin-and-tipping pattern.

  Planned / 60 min
• 16
  

  Waddington Landscapes: Development Follows Constrained Branches

  Maps the framework to developmental biology by adding progressively constrained branching. A cell moves through narrowing channels shaped by gene regulation, where each fork makes alternatives increasingly hard to reverse. Canalization constrains which futures remain reachable. Modern single-cell data connects Waddington's metaphor to measurable gene regulatory attractors.

  Planned / 60 min
• 17
  

  Loss Landscapes: Neural Network Parameters Form the Terrain

  Maps the framework to machine learning: each point is a set of neural network parameters, height is loss (the quantity being minimized, so the goal is a valley rather than a summit), and optimizers navigate a high-dimensional surface of flat minima, sharp minima, and saddle points. The solution found depends on optimizer, initialization, data, and architecture in ways that earlier landscape intuitions help explain.

  Planned / 60 min
• 18
  

  Pareto Landscapes: Multiple Objectives Create a Frontier

  Drops the assumption of a single objective. When multiple goals apply simultaneously, there is no single summit. The Pareto frontier describes the set of nondominated solutions where no objective can be improved without worsening at least one other; away from the frontier, some moves may improve multiple objectives at once. The tradeoff structure matters for engineering, policy, medicine, and AI alignment.

  Planned / 60 min
• 19
  

  Unknown Landscapes: Search Learns a Terrain From Samples

  Drops the assumption that the searcher knows the full value function. In practice, we observe only a tiny sample and must infer the terrain from those observations. Good search balances exploiting what looks promising with exploring regions of high uncertainty. Bayesian optimization, active learning, and ML-guided directed evolution make this concrete.

  Planned / 60 min