Prof. Ofer Shir, Michal Horovitz : Computational Intelligence

We are primarily interested in Operational Research and Machine Learning aspects of experimental domains, that is, algorithmic design for practical physical- and life-sciences problems whose computerized models are unavailable or too expensive to be executed, and thus require real-world measurements when applying search/learning.

We foresee Computational Intelligence algorithms as facilitators that will assist experimental scientists in achieving optimal behavior of their systems and in identifying targeted phenomena. Our long-term research plan is to establish algorithmically-guided discovery tools for bio-systems. A primary pathway, adhering to this plan, is actively to devise specific algorithms for experimental optimization and learning of specific bio-systems.

We are interested in the relations between underlying features of a search landscape to the statistically-learned elements in practice.

Quantum Control Landscapes

We have been interested for some years in mathematical properties of Quantum Control (QC) search landscapes, and particularly when such landscapes undergo global optimization by randomized search heuristics. Such QC landscapes appear in theoretical (QCT) as well as in experimental (QCE) models. We have been studying the capacity to estimate the Hessian matrix over such landscapes, both theoretically in practically (see selected publications in collaboration with Prof. Herschel Rabitz).

Nanduri, A., Shir, O.M., Donovan, A., Ho, T.-S., Rabitz, H.: Exploring the complexity of quantum control optimization trajectories. Physical Chemistry Chemical Physics 17(1) (2015) 334—347

Shir, O.M., Roslund, J., Whitley, D., Rabitz, H.: Efficient retrieval of landscape Hessian: Forced optimal covariance adaptive learning. Physical Review E 89(6) (2014) 063306

Theoretical Statistical Learning of Evolution Strategies

We have been studying in recent years the classical hypothesis of the inverse relation between the statistically-constructed covariance matrix to the landscape Hessian in a randomized search operated by standard Evolution Strategies. Some theoretical results under the quadratic approximation were obtained and published in collaboration with Prof. Amir Yehudayoff, confirming this hypothesis when the population-size tends to infinity.

Shir, O.M., Yehudayoff, A.: On the Statistical Learning Ability of Evolution Strategies. In: Proceedings of the 14th ACM/SIGEVO Conference on Foundations of Genetic Algorithms, FOGA-2017, New York, NY, USA, ACM Press (2017) 127–138

Shir, O.M., Yehudayoff, A.: On the Covariance-Hessian Relation in Evolution Strategies. Theoretical Computer Science 801 (2020) 157—174 DOI: 10.1016/j.tcs.2019.09.002

Experimental science in Chemistry and Biology routinely confronts a shared methodological challenge: identifying optimal combinations of experimental conditions — reagents, concentrations, temperatures, timing, and treatment modalities — within search spaces too vast for exhaustive exploration and too poorly understood for purely model-driven design. The prevailing paradigm of expert-guided, trial-and-error experimentation is both costly and conservative, systematically favouring incremental refinement over principled discovery.
We propose Experimental Combinatorial Optimization (ECO) as a domain-agnostic, AI-driven alternative: an iterative closed-loop framework in which an intelligent search algorithm selects experimental conditions to evaluate, a domain-appropriate assay measures the outcome of interest, and the results guide the algorithm toward progressively superior configurations — without requiring any prior mechanistic model of the underlying chemistry or biology. ECO is thus a conception-free discovery engine: it operates directly on experimental observables, making it applicable wherever a quantifiable assay exists and combinatorial complexity renders exhaustive search infeasible.
The framework has been validated across three domains of increasing biological complexity. In Organic Synthesis, ECO autonomously navigated reaction-condition spaces to identify high-yield protocols. In Protein Expression, it optimized multi-parameter induction and purification conditions without prior structural knowledge [MSc dissertation on algorithmically-guided protein expression]. Most extensively, in Postharvest Technology — the flagship and broadest application to date — ECO has been deployed across multiple commodities, treatment modalities (aqueous treatments, volatile compounds, edible coatings), and regulatory constraints, demonstrating robust performance in both single- and multi-objective settings, including the discovery of consumer-friendly, regulation-compliant protocols from natural-origin substances as principled replacements for synthetic alternatives. [Cucumbers presentation]
Collectively, these results establish ECO as a transferable and extensible methodology for experimental scientists seeking to replace intuition-driven protocols with principled, autonomous discovery — wherever the experimental loop can be closed and the objective can be measured.

A scheme of a typical single iteration (step) in a proposed optimization run of a screening procedure targeting production yield. The input variables for each well are prescribed by the algorithm, whereas the output (feedback) of each produced protein is provided by the assay - altogether closing a feedback loop.

Our long-range goal is to establish algorithmically-guided discovery tools for bio-systems. We foresee Computational Intelligence algorithms as facilitators that will assist experimental scientists in achieving optimal behavior of their systems and in identifying targeted phenomena. A mid-term goal is to devise specific algorithms for experimental combinatorial optimization of bio-systems, and diagnose their capabilities with respect to specific bio-systems' problem-classes.