simulate-stochastic-process
关于
This skill simulates stochastic processes like Markov chains, SDEs, and MCMC samplers, providing sample paths for estimation and prediction. It includes key features like convergence diagnostics, variance reduction, and visualization tools. Use it when analytical solutions are intractable, you need Monte Carlo methods with convergence guarantees, or must sample complex posteriors.
快速安装
Claude Code
推荐npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/simulate-stochastic-process在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Simulate Stochastic Process
Sample paths from stochastic processes — discrete Markov, continuous-time, SDEs, MCMC samplers — w/ convergence diagnostics, variance reduction, trajectory viz.
Use When
- Generate sample paths for est/predict/viz
- Analytical intractable, sim only feasible
- MC est needing convergence guarantees + uncertainty quant
- Validate analytical (stationary, hitting times) vs empirical
- Sample complex posterior via MCMC
- Prototype stochastic model before full analytical
In
Required
| Input | Type | Description |
|---|---|---|
process_type | string | "dtmc", "ctmc", "random_walk", "brownian_motion", "sde", "mcmc" |
parameters | dict | Process-specific (transition matrix, drift/diffusion, target density) |
n_paths | integer | Independent paths to sim |
n_steps | integer | Time steps per path (or total MCMC iters) |
Optional
| Input | Type | Default | Description |
|---|---|---|---|
initial_state | scalar/vector | process-specific | Start state |
dt | float | 0.01 | Time step → continuous discretization |
seed | integer | random | Reproducibility |
burn_in | integer | n_steps / 10 | Initial discard (MCMC) |
thinning | integer | 1 | Keep every k-th → reduce autocorr |
variance_reduction | string | "none" | "none", "antithetic", "stratified", "control_variate" |
target_function | callable | none | Eval along paths → MC est |
Do
Step 1: Define Process + Params
1.1. ID process type + gather params:
- DTMC: Transition matrix
P+ state space. Validate row-stochastic. - CTMC: Rate matrix
Q. Rows sum 0, off-diag non-neg. - Random walk: Step distrib (e.g.
{-1, +1}equal prob), boundaries. - Brownian: Drift
mu, volsigma, dimd. - SDE (Ito): Drift
a(x,t), diffusionb(x,t). - MCMC: Target log-density, proposal (RW Metropolis, HMC, Gibbs).
1.2. Validate consistency:
- Matrix dims match state space size
- SDE coefs satisfy growth + Lipschitz (informal min) for solver
- MCMC proposal well-defined for target support
1.3. Set seed → reproducibility.
Got: Fully spec'd model w/ validated params + reproducible RNG state.
If err: Inconsistent params (e.g. non-stochastic matrix) → correct first. Pathological SDE coefs → diff discretization scheme.
Step 2: Select Sim Method
2.1. Choose algo per type:
| Process | Method | Key Property |
|---|---|---|
| DTMC | Direct sampling from transition row | Exact |
| CTMC | Gillespie algorithm (SSA) | Exact, event-driven |
| CTMC (approx.) | Tau-leaping | Approximate, faster for high rates |
| Random walk | Direct sampling of increments | Exact |
| Brownian motion | Cumulative sum of Gaussian increments | Exact for fixed dt |
| SDE (general) | Euler-Maruyama | Order 0.5 strong, order 1.0 weak |
| SDE (higher order) | Milstein | Order 1.0 strong (scalar noise) |
| SDE (stiff) | Implicit Euler-Maruyama | Stable for stiff drift |
| MCMC (general) | Metropolis-Hastings | Asymptotically exact |
| MCMC (gradient) | Hamiltonian Monte Carlo (HMC) | Better mixing for high dimensions |
| MCMC (conditional) | Gibbs sampler | Exact conditionals when available |
2.2. SDE → dt small enough for stability. Heuristic: start dt = 0.01, halve until results stabilize.
2.3. MCMC → tune proposal scale → acceptance ~:
- 23.4% → high-dim RW Metropolis
- 57.4% → 1D targets
- 65-90% → HMC (depends on trajectory length)
2.4. Variance reduction config:
- Antithetic: Each path w/
Z→ also sim w/-Z - Stratified: Partition prob space, sample within strata
- Control variates: Correlated quantity w/ known E → reduces var
Got: Algo matched to type w/ tuning params.
If err: Unstable (Euler-Maruyama diverging) → implicit method | reduce dt.
Step 3: Implement + Run
3.1. Allocate storage n_paths × n_steps (or dynamic for event-driven Gillespie).
3.2. Per path i = 1, ..., n_paths:
DTMC / Random Walk:
x[0] = initial_state- For
t = 1..n_steps: samplex[t]from transition givenx[t-1]
CTMC (Gillespie):
x[0] = initial_state,time = 0- While
time < T_max:- Total rate
lambda = -Q[x, x] - Holding time
tau ~ Exp(lambda) - Next state from probs
Q[x, j] / lambdaforj != x time += tau, record
- Total rate
SDE (Euler-Maruyama):
x[0] = initial_state- For
t = 1..n_steps:dW = sqrt(dt) * N(0, I)(Wiener)x[t] = x[t-1] + a(x[t-1], t*dt) * dt + b(x[t-1], t*dt) * dW
MCMC (Metropolis-Hastings):
x[0] = initial_state- For
t = 1..n_steps:- Propose
x' ~ q(x' | x[t-1]) alpha = min(1, p(x') * q(x[t-1]|x') / (p(x[t-1]) * q(x'|x[t-1])))- Accept w/ prob
alpha:x[t] = x'if accepted, elsex[t-1] - Record decision
- Propose
3.3. target_function provided → eval at each state, store.
3.4. Apply thinning: keep every thinning-th.
3.5. Discard burn_in from start (MCMC).
Got: n_paths complete trajectories in mem, optional fn evals. MCMC acceptance in target range.
If err: NaN/Inf → reduce dt (SDE) | check params. MCMC accept ~0% | ~100% → adjust proposal scale.
Step 4: Convergence Diagnostics
4.1. Trace plots: per-component over time, subset paths. Visual check stationarity (no trends, stable var).
4.2. Gelman-Rubin (R-hat) for multi-chain MCMC:
- Within-chain
W, between-chainB R_hat = sqrt((n-1)/n + B/(n*W))R_hat < 1.01(strict) |< 1.1(lenient) → convergence
4.3. Effective sample size (ESS):
- Estimate autocorr at increasing lags
ESS = n_samples / (1 + 2 * sum(autocorr))- Rule:
ESS > 400for reliable posterior summaries
4.4. Geweke: cmp mean first 10% vs last 50%. Z-score in [-2, 2] → convergence.
4.5. Non-MCMC: time-avg stats (mean, var) stabilize as path length ↑. Plot running averages.
4.6. Summary table:
| Diagnostic | Value | Threshold | Status |
|---|---|---|---|
| R-hat (max) | ... | < 1.01 | ... |
| ESS (min) | ... | > 400 | ... |
| Geweke z (max abs) | ... | < 2.0 | ... |
| Acceptance rate | ... | 0.15-0.50 | ... |
Got: All diagnostics pass thresholds. Trace shows stable, well-mixing chains.
If err: R-hat > 1.1 → run longer | improve proposal. ESS very low → ↑ thinning | better sampler (HMC). Geweke fails → extend burn-in.
Step 5: Summary Stats + CIs
5.1. Per quantity (state occupancy, fn E, hitting times):
- Point est = sample mean across paths (post burn-in + thin)
- SE via ESS:
SE = SD / sqrt(ESS)
5.2. Build CIs:
- Normal approx:
est +/- z_{alpha/2} * SE - Skewed → percentile bootstrap | batch means
5.3. Variance reduction → VRF:
VRF = Var(naive) / Var(reduced)- Report effective speedup
5.4. MC integration: report est, SE, 95% CI, ESS, # fn evals.
5.5. Distribution est:
- Empirical quantiles (median, 2.5th, 97.5th)
- KDE for continuous
5.6. Tabulate all w/ uncertainties.
Got: Point ests + SEs + CIs. Variance reduction (if applied) → VRF > 1.
If err: CIs too wide → ↑ n_paths | n_steps. Var reduction worsens (VRF < 1) → disable; control variate | antithetic mismatched.
Step 6: Visualize
6.1. Trajectory plots: 5-20 paths over time. Use transparency for overlap.
6.2. Ensemble stats: overlay mean + pointwise 95% CI bands across paths.
6.3. Marginal distributions: at selected times, hist | density estimates of state across paths.
6.4. Stationary cmp: analytical avail → overlay on empirical hist (final time slice).
6.5. Autocorr plots (MCMC): ACF per component, reasonable lag.
6.6. Diagnostic dashboard: trace + ACF + running mean + marginal density → multi-panel.
6.7. Save figures vector (PDF/SVG) + raster (PNG) → docs.
Got: Pub-quality figures showing trajectory, distributional convergence, diagnostics. Analytical (where avail) matches empirical.
If err: Viz reveals non-stationarity | unexpected multimodality → revisit Steps 1-2 (param/method err). Cluttered plots → reduce paths shown | bigger figure.
Check
- All trajectories in valid state space (no out-of-bounds, no NaN/Inf)
- DTMC/CTMC: empirical stationary → analytical (within MC err)
- SDE: halving
dtdoesn't qualitatively change → convergence order - MCMC: R-hat < 1.01, ESS > 400, Geweke z in [-2, 2]
- CI widths shrink ∝
1/sqrt(n_paths)(CLT) - Variance reduction → VRF > 1 (improves not worsens)
- Reproducibility: same seed → identical results
Traps
- Insufficient burn-in (MCMC): Poor initial state → long burn-in before samples represent target. Inspect trace + diagnostics, don't guess.
- Euler-Maruyama instability (stiff SDE): Large drift gradients → explicit can diverge. Implicit | adaptive step.
- Strong vs weak convergence (SDE): Strong = pathwise err (individual trajectories); weak = distributional (expectations). Euler-Maruyama: weak 1.0, strong 0.5.
- PRNG quality: Long sims → low-quality RNGs → correlated samples. Mersenne Twister | PCG | Xoshiro. Verify independence.
- Ignore autocorr (MCMC): Treating autocorr samples as independent underestimates uncertainty. Use ESS, not raw count.
- Antithetic for non-monotone fns: Reduces var only for monotone fn of underlying uniforms. Non-monotone → can ↑ var.
- Mem for large sims: All time steps of many long paths → mem exhaust. Online stats (running mean, var) when full trajectories not needed for viz.
→
- Model Markov Chain — transition matrices + analytical sims validate
- Fit Hidden Markov Model — sim from fitted HMMs → posterior predictive checking + synthetic data
GitHub 仓库
相关推荐技能
llamaguard
其他LlamaGuard是Meta推出的7-8B参数内容审核模型,专门用于过滤LLM的输入和输出内容。它能检测六大安全风险类别(暴力/仇恨、性内容、武器、违禁品、自残、犯罪计划),准确率达94-95%。开发者可通过HuggingFace、vLLM或Sagemaker快速部署,并能与NeMo Guardrails集成实现自动化安全防护。
cost-optimization
其他这个Claude Skill帮助开发者优化云成本,通过资源调整、标记策略和预留实例来降低AWS、Azure和GCP的开支。它适用于减少云支出、分析基础设施成本或实施成本治理策略的场景。关键功能包括提供成本可视化、资源规模调整指导和定价模型优化建议。
quantizing-models-bitsandbytes
其他这个Skill使用bitsandbytes库量化大语言模型,能在GPU内存有限时通过8位或4位量化减少50-75%内存占用,同时保持精度损失最小。它支持INT8、NF4、FP4等多种量化格式,可与HuggingFace Transformers无缝集成,适用于需要部署更大模型或加速推理的场景。还提供QLoRA训练和8位优化器支持,让开发者能轻松实现高效模型压缩。
dispatching-parallel-agents
其他该Skill用于并行处理3个以上无依赖关系的独立故障,可为每个问题域分派专属Claude代理同时执行调查修复。它通过并发处理多个独立问题显著提升故障排查效率,特别适用于测试文件、子系统等无共享状态的场景。
