In recent years, agent based simulation has been successfully utilized in studying various economic phenomena. Those research results focus on modeling and simulation of emotions in Agent Based Economics (ACE).
The economic phenomena simulation based on multi-agent systems is one of the most contemporary ways of studying the latter, as it allows the design of various factors with twenty-two standardized emotions, which interact with each other.
The research scenario under investigation in the context of the present work, is the one of “bank panic”. The term “panic”, refers to the presence of emotional bias towards the unanticipated deposits withdrawal from more than one bank (simultaneous bank runs). The simulation world is composed of depositor agents, influencer agents, banks, ATMs, and retail shops.
What is an agent?
The concept of the agent, first appeared in the 1970s. It was Carl Hewitt in 1977 who developed the Actor model and coined the term. Hewitt proposed the concept of an agent as a self-contained, interactive and concurrently-executing program.
This program had encapsulated internal state, and could respond to messages from other similar programs. The full term of the actor was “Actor is a computational agent which has an email address and a behavior. Actors communicate by message-passing and carry out their actions concurrently”. Additionally, according to the Concise Oxford English Dictionary an agent is: “one who or that which exerts power or produces an effect”.
Agents didn’t appear out of thin air, nor the void. Through the domains of Distributed Artificial Intelligence (DAI), Distributed Problem Solving (DPS) and Parallel AI (PAI) multi-agent systems (MAS) were evolved and through MAS software agents came up. Software agents inherited many of DAI’s motivations, goals and potential benefits.
Generally, an agent could be described as an entity that influences or changes its surrounding. In 1995 Gilbert stated that agents have a degree of autonomy and authority vested. This means that in order to meet their goals/objectives, agents must be allowed to operate, without any kind of direct intervention of humans and should have control of their own actions and internal state.
Also in 1995, in their book, Russell and Norvig gave their own term: “the notion of an agent is meant to be a tool for analyzing systems, not an absolute characterization that divides the world into agents and non-agents”. Based on their term, “agent” can be better described as an umbrella term. So, agents assert in the execution of their own methods. The only exception that occurs, is when another agent requests in accordance with its design objectives, the response agent will execute the request. The requesting agent also has no control over the execution of the agent’s methods. This authority resides within the agent itself.
Agents come with different traits. They can be classified by their ability to move around some network. This yields the classes of static or mobile agents. They can be also classified as either deliberate or reactive.
Agent Based Economics and Agent Based Computational Economics
Agents based simulations and economics, a chronology
Τhe goal of agent-based economics is to create simulations consisting of relatively simple agents that collectively exhibit rich behavior with the overall outcome naturally emerging as a result of their interactions. Agent-based modeling enables the development of macroeconomic models using a bottom up approach.
Each agent is stateful and interacting, and all of them collectively behave in ways that cannot be directly attributed to the decisions of individuals, making the whole greater than the sum of its parts, following the classic complex systems approach. For example, simulated ants colonies can exhibit complex collective behavior despite each individual ant only following trivial rules.
Agent-based models in general, allow us to handle high complexity but they also suffer from a number of shortcomings and restrictions. Individual agents can be programmed and tested one at a time, each of them acting on its own local information and beliefs, but time is directional, many quantities discreet, and aggregate variables cannot be exogenously set. However, one of the primary challenges is quite often their chaotic nature. Allegedly unaccented details and subtle programming errors can decidedly shape the aggregate outcome.
Conventional economic theory, based on mathematics in general and real analysis in particular, begins with a set of definitions and assumptions, and proves theorems. For a number of economists, the dissatisfaction they have with the simple models, is the main reason that led them to the computational approaches. They have been moved into a variety of directions. For example, in public finance, economists sometimes use computation to avoid the single-sector, representative agent models that are commonly used only because of their tractability.
In agent-based computational models, the computer is defined as a catalyst of a mentality change on financial markets. This time it is helping to pursue a world view in which agents may differ in many ways, not just in their information, but in their ability to analyze and process any information, their behaviors toward risks, and in many other dimensions.
Agent based computational economics
Agent based computational economics intend to describe a complex system. A system is defined as complex, if it has the following two properties [see, e.g., Flake (1998)]:
• The system is composed of interacting units;
• The system exhibits emergent properties, that is, properties arising from the interactions of the units that are not properties of the individual units themselves.
An ACE modeler specifies the initial state of an economic system by specifying each agent’s initial data and behavioral methods. An agent’s data might include its type attribute (e.g., world, market, firm, consumer), its structural attributes (e.g., geography, design, cost function, utility function), and information about the other agents attributes (e.g., addresses).
An agent’s methods can include socially instituted behavioral methods (e.g., antitrust laws, market protocols) as well as private methods. Examples of the latter include production and pricing strategies, learning algorithms for updating strategies, and methods for changing methods (e.g., methods for switching from one learning algorithm to another).ACE can be used to a large range of economic systems, no matter if they are micro or macro.
As Camerer (2003) describes, an agent that learns, changes its behavior based on previous experience; and this learning can be calibrated to what actual people are observed to do in the real-world. Additionally , based on Gintis (2000) work, this has as a result, the blend of evolutionary game theory with cultural evolution, the beliefs, preferences, behaviors, and interaction patterns of the agents can vary over time.
An emotion is a feeling, a type of sensation of our brain. we can see the color yellow and perceive the feeling of hunger but neither of the above is an emotion. That being said, an emotion is a sensation of our mind, not of our body. The individual has the final say on his/ her emotions. For example, to be sad means that one feels that he / she is sad. Clore, Ortony and Foss stated in 1987 that if in a sentence the verb “be” can substitute the verb “feel” then the described feeling is an emotion.
For instance, to be sad and to feel sad means the same thing. But feeling neglected is different than being neglected. This test (the “be/ feel” test) usually gives an acceptable answer as to whether or not a mental state is an emotion, but there are some certain exceptions e.g. surprise, lust etc. If an emotion is constant it is then called a “mood”. Emotions on the other hand are meant to be short, instant states of mind the individual passively experiences.
According to the psychological studies, emotions that influence the deliberation and practical reasoning of an agent are characterized as heuristics. That is to prevent excessive deliberation (Damasio 1994. Meyer & Dastani (2004; 2006) proposed a way to map the emotions role in practical reasoning. Based on this approach, an agent is supposed to execute any domain actions to achieve its goals.
The effects of these domain actions cause and/or influence the appraisal of emotions according to a human-inspired model. Sequentially, these emotions influence the deliberation operations of the agent, functioning as heuristics for determining which domain actions have to be chosen next, which completes the circle.
Agents have been involved also in this direction by developing Emotional-BDI agents, that are BDI agents whose behaviour is guided not only by beliefs, desires and intentions, but also by the role of emotions in reasoning and decision-making. They implement emotions such as fear, anxiety and self-confidence. Rao & Georgeff’s BDI logics and Meyer’s KARO framework are built in a way that they don’t allow a simple emotions representation. But, both of them, include properties which can combine in order to properly model Emotional-BDI agents.
Emotions affect an agent’s goals, hence affecting their actions. Emotional effects on goals means reordering existing goals or introducing totally new goals. Emotional states are affected by the goals’ success or failure. An agent which experiences a goal failure may feel unhappy while one experiencing goal success may feel glad.
The Cyprus case
A concrete example of Agent Based Economics is the Cyprus financial crisis scenario. The crisis advanced in the wake of the Greek debt crisis, when Cypriot banks were forced to haircuts of up to 50% in 2011 and the state was unable to raise liquidity from the markets to support its financial sector (Stavárek, 2013). In March 2013 bonds issued by Cyprus were downgraded to Junk status, which disqualified them from being accepted as collateral at the European Central Bank (Wilson, 2012).
Consequently the Cypriot government requested financial aid from the European Financial Stability Facility (EFSF) (Al Jazeera, 2013). The EU-IMF originally proposed to resolve the banking crisis by a bail-out of a €10 billion loan. The results of that political turmoil led to the plan’s alteration, in a way that only depositors at the failing bank were forced to participate in the deal. This was formed as a balance sheet restructuring, a debt-to-equity conversion conversion.
This type of crisis resolution was formally the first realization of a so-called bail-in (Ötker-Robe et al., 2011). The effecting of financial crisis resolution mechanisms has to be evaluated not only by ensuring financial stability, but also by how they impact the entire economy in terms of unemployment, economic growth, liquidity provision to entrepreneurs, etc.
The model that was used for the Cyprus crisis was the Mark I CRISIS model. It consists of a coupled economic and financial Agent Based Modeling, which is closed, i.e. there are no inflows and outflows of any kind of capital. Banking crises can not just be resolved by simply printing money, some agents have to pay for the losses. The model includes three types of agents: households, banks, and firms. Households and firms interact on the labor and consumption-good market, banks and firms interact on the credit market, banks interact on the interbank market.
Βank runs scenario modeling
A bank run is defined as the situation where depositors withdraw their bank deposits because of fear of the safety of their deposits. The term bank panic is quite related to the distortion in reasoning and decision making (emotional state shifting to fear and panic due to rumors spreading) towards a sudden simultaneous withdrawal of deposits from different financial institutions, that lead to parallel bank runs.
Through time, there have been quite some bank runs, such as during the Great Depression in the US. During the 2007 global financial crisis, there were a lot of bank runs internationally recorded (e.g., Countrywide Bank, IndyMac Bank, Northern Rock Bank, etc.).
The selected modelling approach for the Bank Runs scenario was emotions X-Machines.
Samuel Eilenberg introduced in 1974, a variation of Finite State Machines, by claiming it to be a very efficient tool for studying formal languages of the Chomsky hierarchy. They are referred to as X-machines. Many variants have been developed in a number of scientific domains, different from formal languages .
Eilenberg machines define a general computational model. On a given abstract data set X, , a machine is defined as “an automaton labelled with binary relations on X”. The “X” in “X–machine” is a type variable; “an X– machine is a machine or device for manipulating objects of type X”.
A simple example that provides a simple explanation of what X-Machines is, is a calculator that could be described as a float– machine, since it lets us manipulate floating point numbers, and a lorry might be described as a location–machine, because it lets us manipulate where its cargo is located.
A formal model of emotions
For the thesis, emotions are defined as “passions—as defined as event-instigated or object-instigated states of action readiness with control precedence” and the corresponding mode is applied.
These states mirror an agent’s promptitude to choose whether to keep or alter the way it relates to the simulation world. Activation states’ only aim is whether to relate or to not relate (e.g.apathy/unfocused receptivity). Action tendencies are states of readiness related to welcoming or turning away from another agent, situation or event.
Emotion may be contagious. Agents tend to mimic others and this process is influenced by a range of psychophysiological, behavioral, cognitive and emotional factors. Emotional contagion notably depends on emotional perception of others, and the way this perception ends up in imitating their expression.
As perceptions depend on personality, mood and emotion, emotional contagion may be defined as the results of the link between particular personality traits, the mood they will cause on a relentless basis and therefore the intensity of emotion that they will trigger, looking at the intensity of the stimulus.
Emotional X-Machines, a formal model of emotional agents
One of the main characteristics of the X-machines model, is that it has a memory structure. This allows the transitions to be triggered not just by inputs, but by functions that accept an input and the memory values that produce an output and new memory values.
The aforementioned emotional model is fused into an X-Machine model, and thus we have an Emotional X-Machines model that involves emotional states, moods, personality traits and a contagion mechanism. One the main benefits in embracing the Emotional X-Machines methodology, because of its state based approach, is that a feasible and executable model can be derived in a convenient manner.
An agent acting under emotions behaves differently compared to the same agent acting in a rational, emotionless we could say, way. A strong example are cases of disaster management (e.g. emergency evacuation) or bank runs. In such phenomena, agents, according to their personality, face a high probability to experience fear, and finally panic.
For example, in the following figure, an extended, with the emotional structure, model is depicted. Under certain emotional states (e.g. panic due to rumors of financial crisis), the behavior which should be triggered is “withdraw all cash” and not “withdraw some cash”. Such emotional shifting could alter what they perceive and what they communicate to other agents, plus there is also emotional contagion.
Bank Run Modeling in X-Machines
By combining Emotional X-Machines, Emotional contagion with NetLogo, the in discussion, two dimensional world, model, studies bank runs phenomena and their correlation with rumors spreading.
The model focuses on the cash flow and not in electronic transactions (i.e. credit card transactions). This, in purpose, limitation of the model, has made it easier for us to focus on bank panic, i.e. a significant amount of banks failing, a problem that can manifest when depositors withdraw cash for safe keeping at their homes.
The simulation consists of the following agents:
- General population, who are the depositors = white color
- Government agents = blue color
- Journalists = yellow color
and the following infrastructures:
- Companies that the general population works for
Below we have a gif with the simulation.
The blue line is the banks’ cash reserves. The orange one is how many cash have in their wallet, and the gray one, hu much money they desired to have in their wallets.
Here we see the population to get into panic mode really fast, and thus leads to a bank run in the first 117 simulation steps.
Github repo currently private, till further notice.