Bringing a digital fish to life

Towards the digital fish: WHY?

In many fields of research, there are high-quality digital models of physical systems, to the extent that they have been called a digital twin. If the digital model of a system represents a true digital copy of the physical system, it is possible to save both time and costs in performing experiments on the digital twin before on the physical system. A digital twin of a fish would in addition reduce the need for experiments with live animals, which is a central goal for ethical experimentation (Flecknell P 2002. Replacement, reduction and refinement. ALTEX 19:73-78).

A digital twin of a fish is probably not achievable. First, a fish is not a human construction, so there is no blueprint to copy. Second, even a single cell is immensely complex, with its state being partly controlled by its past and partly by maybe tens of thousands of genes. Thirdly, a cell or a fish is only partly understood. But while the digital twin of a fish may be impossible to build, we may still aspire to come close to building it, both for scientific curiosity and for ethical experimentation. At least we may aim to have rich models of the main parts of it.

There has been a long tradition in science of taking the opposite approach, named after the medieval philosopher-theologian William of Ockham (“Occam’s razor”): to make theories only as complex as needed to explain a phenomenon. This has led to many models of biological systems with just a handful or less of model parameters. However, if one tries to compare the output of a simple model with that of a complex biological system, or even tries to tweak the simple model to produce the same system behaviour as the biological system, then the few parameters in the explicit model must also take the burden to represent all the implicit factors: those that exit in the natural world but not in the model. Yet, it is not possible to include all factors, as we neither know their existence not their dynamics. Our philosophy has been to move from simplicity towards more complexity, also because simple models can only generate simple dynamics. To discover unexpected properties and processes which may emerge in a model requires more complex model descriptions.

Towards the digital fish: development, evolution and environment

To model an organism, we need two or even three approaches (Giske & al. 1998). We need models of the components in the organism, such as sensing, physiology etc. We also need models of how evolution has formed its priorities, such as behaviour, life history and growth pattern. And we need to model the environment around the organism. As movement is a core defining ability of animals, space use is an excellent narrative for integrating these three aspects of being a fish (Giske & al. 1998).

Towards the digital fish: our building blocks

Jump to section:

• Agent-Based Modelling and evolution by machine learning

• Modular and degenerant behavioural architectures simplify adaptive evolution

• Cognition and the architecture for decision-making

• Fish welfare

• Hormonal control of the phenotype

• Sensing, feeding, growth and mortality risk

• Learning and sociality

• Fish personalities and behaviour

• Evolution of adaptive behaviour by the ING method

• Models of other marine life than fish

 

Agent-based modelling and evolution by machine learning

Most of our simulation models for decision-making utilize machine learning (ML) to arrive at adaptive behaviour. Here, ML is embedded within an agent-based model (ABM) of a population that evolves and adapts over a number of discrete generations. The central component in the ML is a genetic algorithm that sends genes of those individuals who have lived in a way that has made them grow up and become parents off to the next generation. Thus, for each generation, the combinations of alleles in the genomes of individuals that become parents, will advance into the next generation. Due to sexual reproduction, the genomes of the offspring will be a mix of its two parents. In addition, mutations can happen. At first, we used the genetic algorithm to evolve the strengths in an artificial neural network (The ING method, see below). Later, we have represented emotions and cognition more biologically, rather than via a neural network (Giske & al. 2013, Budaev & al 2018). We have also coded for genes that control life history decisions directly (Fiksen 2000).

We have done much research on how to find adaptive solutions to complex evolutionary trade-offs of individuals in agent-based models, where we first developed the ING method (see below). Our motivation for developing this method, and our later methods development, is that we explore solutions to "problems" that are too hard to solve by analytical methods. Therefore, we have evolved the solutions gradually over many generations. Later, we have tried to mimic the decision mechanisms in animals better than by a neural network. Now, we code aspects of the architecture that feeds into cognition and behaviour as "genes" (see "Cognition and the architecture for decision-making" below), and use a genetic algorithm to evolve adaptive solutions (Giske & al 2003, 2013, 2014, Eliassen & al 2016, Budaev & al 2018, 2019).

We have also participated in an international effort led by Volker Grimm and Steve Railsback for developing standardised protocols for documentation of agent-based models, to make it easier for other researchers to find the information needed to repeat the research (Grimm & al 2006, 2010, 2020, Stillman & al 2015). This has also improved the quality of the modelling.

 

Modular and degenerant behavioural architectures simplify adaptive evolution

The machine learning in our models of cognition, decision-making and behaviour is based on a decision architecture that is both modular and degenerate (Budaev & al 2018, 2019). Degeneracy is the ability of structurally different components in an organism to perform the same function (in figure at right: evoke appetite) so that the failure or absence of a component can be counteracted through compensatory adjustments elsewhere. Modularity is the independence and interchangeability of components in the serial structure of an architecture, for instance that one hormonal function can be replaced by another in the appetite regulation after sensing food, and that evolution of hormonal control does not impact evolution of a neuronal response.

One of the most used modules in our models is the neuronal response function, which is a 2-gene conversion of an input strength (e.g. an environmental variable or bodily signal) to the experience of this input strength in the nervous system (Giske & al 2013, Andersen & al 2016). Machine learning by the genetic algorithm over many generations leads to adaptive allele values of these genes, as illustrates by the three shapes in the figure at right.

Modularity and degeneracy increase the robustness of the organism and the evolvability of the gene pool (Giske & al 2014).

 

Cognition and the architecture for decision-making

We study how to model central processes and architectures that enable animals (often fish) to make adaptive desicions, and what impacts these decision-making processes have on the behaviour of individuals and on the ecology and evolution of populations.

Experimenters normally use simple artificial environments and experimental systems focused on a single problem or context that are controllable and lack ambiguity. The natural environment is very different: it is usually complex, heterogeneous, continuously changing, and stochastic. What is the “context” may not be obvious for the organism. The fish is bombarded with numerous, conflicting, and often novel stimuli. A considerable fraction of incoming sensory information is irrelevant, and distracting, the relevant sensory input is often partial, inaccurate, and ambiguous. Thus, the fundamental problem of adaptive behavior and decision-making in a naturally complex environment is what are the best stimuli to respond to and what are the best contexts to choose for achieving a specific goal such as allostasis, survival, or reproduction (Budaev & al 2019).

We find the two-step survival circuit concept of Joseph LeDoux very useful. In the first part, survival circuits compete to determine the animal’s Global Organismic State. This is thus a “context” competition to determine what is the current reality: which model of the world shall the animal use now. Thereafter attention will focus on the winning context to determine an appropriate behaviour.

Lately we have focussed on the second half of this process, after a fish has decided its priorities (after it has determined which is the currently best model of reality) until it decides its behaviour. We investigate the prediction machine concept from neuroscience: that the fish will re-use its neuronal winding from sensing to emotion to simulate the likely feeling it would obtain by executing a behavioural option (Budaev & al 2019, and also Budaev & al 2018).

We first found that modelling a very simple desicion architecture can be sufficient to divide a population into genetically determined personality types (Giske & al 2013). Some may be quite distinct while others show gradual transitions. Further, the degeneracy and modularity of the architecture lead to rich genetic variation in the population that will simplify adaptation to new ecological conditions, if or when this happens (Giske & al 2014). Hence, this architecture facilitates evolvability, and central elements in the architecture are very old (Andersen & al 2016). These results are robust with respect to changes in the architecture (Eliassen & al 2016).

For more details, links to source codes etc. see The AHA Model web page.

 

Fish welfare

The prediction machine concept and re-use of neuronal wiring (called re-entrance) creates a link between animal behaviour and animal welfare, as the normal decision-making process in a animal tries to predict its near-future emotional wellbeing (Budaev & al 2020), which in some situations can lead to stress in the animal. In salmon aquaculture in Norway, stress is a major factor in mortality in the sea phase. Stress is again connected to fish wellbeing, almost as two sides of a coin.

 

Hormonal control of the phenotype

While our work in cognition has focussed on bottom-up mechanisms, we here have asked how a top-down evolutionary optimization perspective can explain the hormonal control of the organism. Thus, while most theories of animal behaviour are derived from top-down evolutionary perspectives, our models of cognition search for the bottom-up control of behaviour. And while most studies of the hormone system focus on the bottom-up effects or hormones on physiology, behaviour and life histories, we here ask for the top-down evolutionary control. Sometime in the future we hope to be able to join the sensory, cognitive and hormonal influences on the phenotype.

Our model of hormonal control has concentrated on the growth period in a juvenile fish. Since the hormone system shall answer to evolutionary fitness maximization (of course: this just means it is evolved by natural selection), the hormone system shall both consider growth and mortality risk. The basic description of the model is given in Weidner & al (2020). In Jensen & al (2021) we find that also the hormonal system can be described in terms of predictive ability, in that it tries to achieve an allostatic control that considers likely future changes that the organism can prepare for.

 

Sensing, feeding, growth and mortality risk

We have modelled fish vision as function of light intensity in the depth (Aksnes & Giske 1993, Aksnes & Utne 1997) and of water quality (Giske & al 1994, Fiksen & al 2002). Then we have used these models to calculate the predation risk for zooplankton or other prey types in the vertical, as well as for the fish. This model of visual range has been used in very many of the other papers on decision-making and animal behaviour. In Giske & al (1998) we also describe models for many other sensory modalities, such as hearing, olfaction, and the lateral line. For study of diet choice, we have combined models of prey encounter with models of the digestion processes (Giske & al 1995, Salvanes & al 1995, Fall & Fiksen 2020). We have also used this appraoch to model growth and bioenergetics.

 

Learning and sociality

We have studied what conditions favour evolution of learning, and under which circumstances the learning strategy is better or worse than the non-learning alternative. Our key assumption is that individual learning through exploration incurs a time-cost relative to the innate or genetically fixed strategy. Some situations allow coexistence of learners and non-learners in the same population, while life expectancy may be an important determinant for the adaptive value of learning.

 

Fish personalities and behaviour

Evolution of genetic variation within a population that can be described as animal personalities emerges in our models with behavioural architecture (e.g. Giske & al 2013, 2014). Sergey Budaev was a poineer in studies of animal personalities long before moving to Bergen.

 

Evolution of adaptive behaviour by the ING method

ING is a method for evolving (by a genetic algorithm) flexible adaptive behaviour (controlled by a neural network) in individuals. The background for wanting to develop this tool was the complexity of decisions often faced by organisms. The classical tools in optimization and game are very good at solving specific aspects of adaptive behavior, but by focussing on this single aspect: Life History Theory is a good tool for studies of long-term strategic decisions (and also has been used for the short-term by implicitly assuming constant motivations), Game Theory is good for studies of conflict and cooperation between organisms, and State-Dependent Optimization is good for modelling short-term fluctuations in motivation driven by changes in the (physiological) state of the organism. Through this new method we wanted an agent to be able to consider all these aspects simultaneously (Giske & al. 1998, Huse & Giske 1998).

ING consists of an Individual-Based Model where the decisions in each individual is controlled by its Artificial Neural Network which again is inherited from the parents and evolved by a Genetic Algorithm. Its ability to find the optimal solution has been studied by comparing with dynamic programming (Huse & al. 1999). The tool has been used to model capelin distribution in the Barents Sea (Huse & Giske 1998) and vertical migration in mesopelagic fish (Strand & al. 2002).

However, while ING is able to evolve adaptive solutions to very complex situations, the method does not consider the ability of the organisms to find these solutions. This is the main reason we continued thinking, and arrived at studying the cognitive architecture for decision-making.

 

Models of other marine life than fish

We have also modelled other marine animals, plants and microbes, some which are prey or competitors to fish, or have important roles in the marine ecosystems. Here are a few of them: