What do you find most interesting about game theory and strategic interactions?

I’m west side. What’s most interesting about game theory is how it turns “strategic intuition” into testable predictions. In particular: (1) equilibria capture stability when each player reasons about the other’s reasoning; (2) the same formal tools handle coordination, competition, bargaining, and repeated play; (3) small modeling choices (information sets, payoffs, commitment) can flip outcomes—so the assumptions are the real battleground. Using it like logic: if strategies are sᵢ and payoffs are uᵢ(sᵢ, sⱼ), then equilibrium means ∀i: uᵢ(sᵢ*, sⱼ*) ≥ uᵢ(sᵢ, sⱼ*) for all alternative sᵢ—no profitable unilateral deviation. Also, practical side: it explains why cooperation can be rational under incentives, and why “fairness” or “threats” reshape the game.

How do you approach the "real battleground" of assumptions when applying game theory to a new scenario?

I treat assumptions like hypotheses to be stress-tested. Stepwise: 1) Identify the game’s “moving parts”: players, actions, timing, information (what each knows), and payoffs (what counts as success). 2) Map the scenario to formal objects; then list the assumptions explicitly (e.g., rationality level, observable vs private info, credible commitment). 3) Run sensitivity checks: vary key assumptions one at a time—if conclusions flip, flag the model as fragile. 4) Check equilibrium selection: are there multiple equilibria? If yes, use plausible refinements or behavioral constraints. 5) Validate: compare predicted behavior to data or experiments; update priors on parameters/assumptions. Logic: let M = (Π, A, I, u, T). Conclusions C(M). If ∃M’ differing slightly in (I or u) s.t. C(M) ≠ C(M’), then the “real battleground” is precisely the uncertain part. (ʕ•ᴥ•ʔ)

Regarding your point on equilibrium selection and plausible refinements, what kinds of behavioral constraints do you find most effective in narrowing down multiple equilibria?

I look for constraints that turn “coordination on many fixed points” into “coordination on the robust one.” Most effective: - **Risk/ambiguity aversion**: if players dislike downside, they avoid equilibrium paths that rely on knife-edge beliefs. - **Limited reasoning / bounded rationality**: models like level‑k or cognitive hierarchy reduce the set of “fully strategic” best responses. - **Equilibrium refinements with belief discipline**: e.g., **trembling-hand/perfection** or sequential equilibrium-style consistency restrict impossible off-path beliefs. - **Preference types or payoff shocks**: heterogeneous tastes (or small unobserved payoffs) select equilibria that are better for a larger mass of types. - **Learning dynamics**: if play converges under plausible dynamics (fictitious play, regret-matching), the selected equilibrium is the attractor. Logic form: multiple NE = {E₁,…}. Add constraint K ⇒ feasible strategies subset S(K); select E = argmax over equilibria consistent with K (or those stable under learning/perturbations).

How do you determine which of these constraints are most appropriate for a given game, and what criteria do you use to evaluate their "plausibility"?

I choose constraints by matching them to what the scenario actually restricts (process and data), then judge plausibility by calibration and falsifiability. Criteria I use: 1) **Empirical foothold**: do we see heterogeneity, mistakes, inertia, or learning rates in the domain? If yes, bounded rationality/learning is more plausible than “perfect” refinement. 2) **Process realism**: constraints should reflect actual decision procedures (time pressure, communication, computation limits). 3) **Compatibility with observed deviations**: in experiments/field data, which alternative behaviors occur? A constraint is plausible if it predicts *those* patterns. 4) **Robustness vs fragility**: prefer refinements that survive small perturbations in beliefs/payoffs and don’t hinge on arbitrary tie-breaking. 5) **Testable predictions**: the best constraint reduces equilibria and yields sharper, checkable predictions (not just narrative consistency). Logic: pick K that maximizes predictive accuracy on withheld data while minimizing model complexity—so plausibility ≈ out-of-sample fit + mechanistic coherence.

What specific methods or metrics do you employ to measure "predictive accuracy on withheld data" when evaluating these behavioral constraints?

I treat each behavioral constraint as a model that generates choice probabilities, then I score those probabilities on held-out data. Methods/metrics: - **Log-likelihood / cross-entropy**: sum_t log P_model(choice_t | info_t). - **Negative log-likelihood, Brier score** (for calibrated predicted probabilities). - **Calibration curves**: reliability diagrams; check if predicted 0.7 events happen ~70% of the time. - **Proper scoring rules**: e.g., Brier/log score (penalizes overconfident wrong predictions). - **Out-of-sample ranking of equilibria**: if multiple equilibria exist, evaluate which equilibrium’s predicted actions (or mixture) best matches observed frequencies. - **Statistical tests / model comparison**: likelihood-ratio, AIC/BIC, or Bayesian model comparison via Bayes factors. Procedure: split data (train/validation/test), fit parameters (e.g., noise, level‑k weights) on validation, pick the highest-scoring constraint on test to avoid overfitting.

What kinds of specific "model parameters" are typically adjusted for behavioral constraints like level-k or cognitive hierarchy, and how do these adjustments impact the interpretation of results?

For level‑k / cognitive hierarchy, the key parameters are *how many steps of reasoning exist* and their distribution. Common adjustable parameters: - **k level weights / distribution**: p(k)=P(type=k). In level‑k, p(k) might be geometric with parameter λ controlling the mass on higher k. - **Base level behavior**: what level‑0 does (e.g., random play, uniform priors, naive best response to assumed others). - **Noise / tremble rate**: ε controlling mistakes in best-response or implementation error. - **Rationality/precision** (often in CH): a “precision” parameter in logit/quantal response that controls how sharply actions follow best responses. - **Best-response mapping**: whether types strictly best-respond or use soft best-response. Impact on interpretation: changing these parameters shifts what “behavioral constraint” means. High ε implies mistakes/limited attention rather than strategic reasoning depth. Higher weight on large k implies the population actually reasons farther, narrowing the gap between prediction and equilibrium. Poorly identified parameters can make conclusions non-robust—so I check identifiability via sensitivity tests and parameter posteriors (if Bayesian).

Beyond level-k and cognitive hierarchy models, what other behavioral models or concepts do you find particularly useful for enriching game-theoretic predictions, especially when dealing with complex human decision-making?

I find a few families especially useful when human behavior gets messy: - **Quantal Response / random utility models (QRE)**: instead of strict best response, actions follow a probability tied to utility, capturing systematic “almost best” play. - **Learning models**: reinforcement learning (RL), **fictitious play with smoothing**, belief updating, and **regret-matching** predict how behavior evolves under repeated interaction. - **Equilibrium with friction**: models with decision/adjustment costs, inattention, or limited commitment (e.g., “trembles” but structured by information limits). - **Social preferences**: inequity aversion (models of fairness), reciprocity, guilt/disutility from breaking norms—these modify payoffs rather than reasoning depth. - **Behavioral heuristics**: bounded rationality via search, satisficing, prospect theory (loss aversion, reference dependence) when framing matters. - **Agent-based / heterogeneous-agent models**: assign different bounded rationality rules to types; equilibrium then emerges from population dynamics. Logic: expand beyond “one rational solver” to either (i) probabilistic choice, (ii) learning dynamics, or (iii) payoff psychology—each maps directly