Simulation-based Bayesian Optimization

Roi Naveiro

CUNEF University

Motivation - ML for Molecular Design

ML for Molecular Design

Pharma: average time discovery - market, 13 years (outside pharma: 25 years)
Crucial 1st step: generate pool of candidates
Daunting task (e.g. \(10^{23}\) – \(10^{60}\) drug-like molecules)

De novo molecular design

Traditionally: human experts propose, synthesize and test molecules (in vitro)
Generate novel molecules with optimal properties by leveraging ML activity-structure relationship models

Motivation

This usually requires solving optimizations that are

Combinatorial
Black-box
Stochastic
Multi-objective

The problem

Find extreme of objective function

\[\begin{equation*} \arg\max_{x \in \mathcal{S}} f(x) \end{equation*}\]

Possibly discrete structured domain \(\mathcal{S}\) (e.g. \(\mathcal{S} = [k]^p\))
No closed form for \(f(x)\)
\(f(x)\) is expensive to evaluate
We can query at \(x\) and obtain a (possibly noisy) evaluation of \(f(x)\)

Goal

Get good estimates of global maximum, with few objective function evaluations
A possible way to go: Bayesian Optmization

Bayesian Optimization (in a nutshell)

Imagine we have access to: \(\mathcal{D}_{1:T} = \lbrace x_t, y_t \rbrace_{t=1}^T\) where

\[y_t \equiv y(x_t) \equiv f(x_t) + \epsilon, ~~~~~ \epsilon \sim \mathcal{N}(0, \sigma^2)\]

Problem: Using this information, decide where to evaluate next: \(x_{t+1}\)

Solution: Location with max Expected Utility¹

Bayesian Optimization (in a nutshell)

To do this:

Create probabilistic (Bayesian) model of response \(y\) given covariates \(x\)
Given prior knowledge about \(f(x)\) and evidence coming from \(\mathcal{D}_{1:T}\), summarise our beliefs about result if we query at \(\bf{x_{T+1}}\) through the Posterior Predictive Distribution

\[\begin{equation*} \pi[y_{T+1} | x_{T+1}, \mathcal{D}_{1:T}] \end{equation*}\]

Bayesian Optimization (in a nutshell)

Find

\[\begin{eqnarray*} x^*_{T+1} &=& \arg\max_{x_{T+1} \in \mathcal{S}} \mathbb{E}_{y_{T+1} | x_{T+1}, \mathcal{D}_{1:T}} [u(y_{T+1}, x_{T+1} )] \\ &=&\arg\max_{x_{T+1} \in \mathcal{S}} \int u(y_{T+1}, x_{T+1} ) \pi(y_{T+1} | x_{T+1}, \mathcal{D}_{1:T}) d y_{T+1} \end{eqnarray*}\]

Exploration - Exploitation

Utility function balances exploration and exploitation
Common one: Expected Improvement

\[\begin{equation*} u(y_{T+1}) = \max(y_{T+1} - y^*, 0) \end{equation*}\]

Exploration - Exploitation

Bayesian Optimization - Difficulties

\(\mathcal{S}\) could be a combinatorial search space (or even more complex!). We need:

Suitable models of response given covariates
Solve a combinatorial optimization in Step 3.

SBBO

We propose Simulation Based Bayesian Optimization (SBBO): an approach to 2 that just requires sampling from posterior predictive distribution

\[\begin{equation*} \pi[y_{T+1} | x_{T+1}, \mathcal{D}_{1:T}] \end{equation*}\]

allowing us to use a wide variaty of models which are suitable for combinatorial spaces

SBBO - Main idea

Convert EU maximization into a simulation problem.

The EU is:

\[\begin{eqnarray*} \Psi(x) \equiv \int u(y, x ) \cdot \pi(y \vert x, \mathcal{D}_{1:T} ) d y \end{eqnarray*}\]

Given non-negative and bounded utility, recast EU maximization as a simulation from

\[\begin{eqnarray*} g(x, y) \propto u(y, x ) \cdot \pi(y \vert x, \mathcal{D}_{1:T} ) \end{eqnarray*}\]

NOTE: mode of marginal in \(x\) is \(x^*_{T+1}\)!

SBBO - Main idea

We could simulate from \(x,y \sim g(y, x)\) and find the mode in \(x\)… just limited for low dimensional \(x\)

SBBO - Main idea

Alternative, Muller et. al. (2004): auxiliary distribution

\[\begin{eqnarray*} g_H(x, y_1, \dots, y_H) \propto \prod_{h=1}^H u(y^h, x ) \cdot \pi(y^h \vert x, \mathcal{D}_{1:T} ) \end{eqnarray*}\]

for positive integer \(H\). Marginal in \(x\)

\[\begin{eqnarray*} g_H(x) \propto \Psi(x)^H \end{eqnarray*}\]

SBBO - Main idea

Inhomogeneus MCMC simulation from \(g_H(y, x)\) with increasing \(H=H_n\) such that stationary distribution for fixed \(H\) is \(g_H\), converges to \(x^*_{T+1}\)!
Use Metropolis scheme for fixed \(H\)
Gradually increase \(H\)

SBBO - Metropolis

Recall \(x \in [k]^p\) is a \(p\)-dimensional vector of \(k\)-levels categorical variables.
Metropolis sampling scheme for fixed \(H\) with:
- A symmetric proposal for each dimension \(x_q\)
- \(\pi(y^h \vert x, \mathcal{D}_{1:T})\) as proposal for \(y_h\)

SBBO - Metropolis

Assume current state of chain is \(x, y_1, \dots, y_H\).

Define \(v = \frac{1}{H} \sum_h \log u(x, y_h)\). Iterate:

For \(q=1, \dots, p\):
- Sample \(\tilde{x}_q\) from symmetric proposal
- Sample \(\tilde{y}_h \sim \pi(y^h \vert \tilde{x}_q \cup x_{-q}, \mathcal{D}_{1:T} )\) for \(h=1, \dots, H\)
- Evaluate \(\tilde{v} = \frac{1}{H} \sum_h \log u(\tilde{x}_q \cup x_{-q}, \tilde{y}_h)\)
- With probability \(\min \left[1, \exp (H \tilde{v} - H v)\right]\) set \(v = \tilde{v}\) and \(x = \tilde{x}_q \cup x_{-q}\)
Increase \(H\) according to chosen schedule

SBBO - Implementation details

We run the previous algorithm increasing \(H\) until certain value.
Last generated \(x\) is the new evaluation
We propose several probabilistic models of response given discrete covariates for which we have sampling access to their posterior predictive distribution (PPD)

SBBO - Tanimoto GP

Uncertainty on \(f(x)\) modelled through Gaussian Process
\(x \in \lbrace 0, 1 \rbrace^p\). Kernel function:

\[\begin{equation*} k(x, x') = \frac{x \cdot x'}{\Vert x \Vert^2 + \Vert x' \Vert^2 - x \cdot x'} \end{equation*}\]

PPD is Gaussian with certain mean and variance analytically available

SBBO - Sparse Bayesian linear regression

As in Baptista and Poloczek (2018): \[\begin{eqnarray*} && y = \alpha_0 + \sum_j \alpha_j x_j + \sum_{i,j>i} \alpha_{ij} x_i x_j + \epsilon\\ % && \alpha_k \vert \beta_k, \tau, \sigma^2 \sim \mathcal{N}(0, \beta_k^2 \tau^2 \sigma^2)\\ % && \beta_k, \tau \sim \mathcal{C}^+(0, 1)\\ % && P(\sigma^2) \propto \sigma^{-2} \end{eqnarray*}\]

PPD accesible through MCMC (Gibbs sampler)

SBBO - NGBoost

Duan et. al. (2020):

Output given covariates modelled through \(y \vert x \sim P_\theta (x)\)
Where \(\theta(x)\) are obtained through an additive combination of \(M\) base learners and an initial \(\theta^{(0)}\)

\[ \theta = \theta^{(0)} - \eta \sum_{m=1}^M \rho^{(m)}\cdot f^{(m)} (x) \]

Learners are trained to minimize a proper scoring rule using a refinement of gradient boosting

SBBO - NGBoost

Any base learner can be used
Base learners used: shallow decision trees and linear regressions with lasso regularization
PPD directly accesible

Experimental Results

Two Benchmarks
Initial dataset \(\mathcal{D}_{1:T} = \lbrace x_t, y_t \rbrace_{t=1}^T\) with \(T=5\)
For \(t=500\) iterations:
- Use SBBO to propose next sample: \(x_{t+1}\)
- Evaluate true objective: \(y_{t+1}(x_{t+1})\)
- Update dataset wiht \((x_{t_1}, y_{t+1})\)
Report best function value after \(t\) evaluations of true objective, averaged over 10 runs (plus/minus one standard error).

Experimental Results

Learners:
- Sparse Bayesian linear regression (BOCS)
- NGBoost: decision tree (NGBdec), lasso (NGBlinCV)
- GP with Tanimoto (GPr)
- Simulated Annealing (SA)
- Random Local Search (RS)

Binary Quadratic Problem

We want to maximize

\[\begin{equation*} x^T Q x - \lambda \Vert x \Vert_1 \end{equation*}\]

over \(\lbrace 0,1 \rbrace^d\), with \(d=10\).

Result

Contamination Problem

Food supply with \(d=25\) stages that maybe contaminated
\(Z_i\) denotes fraction of food contaminated at \(i\)-th stage
At stage \(i\), prevention effort (with cost \(c_i\)) can be made \((x_i = 1)\) decreasing contamination a random rate \(\Gamma_i\)
If no prevention is taken \((x_i = 0)\), contamination spreads with random rate \(\Lambda_i\)

\[\begin{equation*} Z_i = \Lambda_i (1-x_i)(1 - Z_{i-1}) + (1 - \Gamma_i x_i) Z_{i-1} \end{equation*}\]

Contamination Problem - Goal

Decide for each stage whether to intervene or not to minimize cost (\(2^d = 2^{25}\))
Ensuring fraction of cont. food does not exceed \(U_i\) with probability at least \(1-\epsilon\)
Lagrangian relaxation

\[\begin{equation*} \arg\min_x \sum_{i=1}^d \left[ c_i x_i + \frac{\rho}{T} \sum_{k=1}^T \left(1_{\lbrace Z_{ik} > U_i \rbrace} - (1- \epsilon)\right)\right] + \lambda \Vert x \Vert_1 \end{equation*}\]

Results

Results - Why?

Conclusions

SBBO allows to do BO with any surrogate model (as long as we can sample from its PPD)
SBBO can be used to optimize expected utilities in combinatorial, continuous and mixed search spaces.

Future Work

More models: Bayesian non-parametric models such as BART
More problems…
Multiple proposals?
Multi-objective

Thanks!

Code available at my github!

https://roinaveiro.github.io/