Oliveira

The MAS simulation implemented by Oliveira in [139] aims to analyse how emotions can be used to facilitate the adaptation of an agent’s behaviour to that employed by others and whether or not emotion is capable of enhancing the monetary utility of individual agents. Consequently, the research motivations for this work can be classed as being a part of the cognitive-engineering research motivation proposed in [25]. A public goods game where pieces of a pie may be bid for is simulated (each agent may only bid for a maximum of three pieces) and, if the total bids from all agents are less than or equal to the total number of slices available, each agent receives the number of pieces they bid for. If the total number of pieces bid for by all agents is greater than the total number of pieces available then agents do not receive anything. As can be inferred from the game description, each agent’s utility is equivalent to the number of pie pieces they receive at the end of the game.

The closing function and adaptation functions mentioned above are used in context of an automaton, defined by Oliveira as:

“a decision rule consisting of a finite set of states, a transition function (which defines the transition between states) and a behavioural function (defining how the agent behaves in each state)”.

An automaton therefore maps out the actions an agent can perform in its current state, the actions that may be performed in every possible state given the current state and the set of all states possible if a particular action is performed in a particular state. Oliveira notes that such automatons are difficult to compute if an agent posses limited or incomplete information about its environment (this echoes Simon’s concept of “bounded rationality” [174]).

Consequently, an agent may reach an impasse in its decision-making if the automaton it has computed does not yield either a beneficial state or an action to achieve a beneficial state. In such a case, closing and adaptation functions are used: a closing function is used to close an automaton i.e. move the agent from one state to another whilst an adaptation function is used if an agent’s best response (defined as behaviour that maximises the agent’s individual total of discounted rewards) fails to achieve desirable equilibria. The particular behaviour exhibited when these functions are employed is dependent upon the emotions that the agent is endowed with.

Each agent in Oliveira’s simulation is endowed with a maximum of two, distinct, emotions (which can be anger, apathy or patience) for its closing/adaptation function e.g. an agent may use anger for its closing function and apathy for its adaptation function. Therefore, rather than being modelled as distinct notions in their own right, emotions are modelled as strategies, much like Jiang et al.’s implemented emotional system discussed in section 4.3.5.

Oliveira’s method of modelling emotions is particularly interesting since it does not appear to be based upon any general psychological emotional model. Instead, Oliveira models four emotions based upon research into their mechanisms conducted by indepen- dent parties, these emotions are:

ˆ Anger: the agent attempts to minimize his opponents’ rewards.

ˆ Apathy: the agent always plays the same action independently of the rewards received.

ˆ Patience: the agent attempts to reward his opponents if they change their be- haviour in order to benefit the community.

ˆ Confidence: the agent plays the same and does not adapt according to best re- sponse aiming to convince other player’s to use his strategy; this gains the agent credibility.

In Oliveira’s simulation, emotions are functional in that they directly enable agents to make decisions and perform intentional behaviour in context of competitive, resource- bounded environments. The intentional behaviour produced by these emotions is guar- anteed to occur and may be focused upon both an agent and its opponents. This means that the concept ofaction potentials (see section 4.1.1.2) is again not considered in this emotional model. With regards to emotion-behaviour mapping it would appear that one emotion can be mapped to many behaviours as Oliveira himself stipulates when discussing the effects of the emotions modelled.

There are a few important points to note with respect to Oliveira’s use of emotion in [139]. Firstly, “rational” behaviour is defined by Oliveira as the ability of an agent to calculate its best response in any given automaton. So, if an agent is not capable of calculating its best response then it may use an “emotional” strategy to close the automaton; emotion is used when rationality is ineffective. The implicit issue with respect to this is that Oliveira is proposing a dichotomy between rational and emotional behaviour in much the same way as Jiang et al. do (see section 4.3.5). Furthermore, only endowing an agent with one emotion for use with each function seems to produce quite inflexible behaviour. The agent’s ability to compute closing and adaptation functions could be much more extensible if more than one emotion could be used with respect to the closing or adaptation function. Consequently, it can be asserted that opposing emotions may never be activated simultaneously in an agent since an agent only ever has one emotion available to it and there are no antonyms of the emotions modelled.

Furthermore, like the implemented emotional models discussed in sections 4.3.4 and 4.3.5, Oliveira’s emotional model does not consider the notion ofpotential oractivation thresholds for emotions. Instead, if a best-response cannot be calculated by an agent, the emotion that an agent has been endowed with for its closing function is used. The situation is identical when an agent uses an emotion in context of its adaptation function.

In this way, the intensity of the emotions implemented by Oliveira are constant and do not reflect a human-centric approach toward emotion modelling.