CARACTERÍSTICAS DE LOS GRUPOS SOCIO- SOCIO-ECONÓMICOS (MONTECRISTI)
3.6 FUENTES DE ALIMENTOS
The communication in this thesis is the process of evolving languages and generating interesting design works among agents playing different roles, such as speaker and listener in guessing games, and client and designer in generation games. In each language game,
the agents learn successful relations between topics and utterances using Equation 6
(initial weight: w0 = 0.01, learning rate: η = 0.1) to strengthen their weights, and
unlearn failed relations using Equation7 to weaken their weights.
w ← η + (1 − η) × w (6)
w ← (1 − η) × w (7)
In each language game, weighted random choice (i.e., weighted choice using roulette
wheel selection p(x) = wx/Pni=1wi) of utterances or designs is typically better than
either “best” choice (i.e., choosing the item with the highest weight) or random choice (i.e., equal probability of choosing any item). Compared with “best” choice, weighted random choice enables an item with low weight still having a small chance of being selected; compared with random choice, it provides more reasonable results.
Whether to end or continue language games is determined by success rate (e.g. when ratesuccess > 0.7, the game will stop and new type of game will start) and the number
of games being played (e.g. nmin = 100, nmax = 1000, nmin <= n <= nmax. Setting
nmin is to avoid premature success). The success rate is calculated using Equation 8.
rs= ns/n (8)
rs is the current success rate, ns is the number of successful games while n is the total
number of games played by agents.
The processes of communication involve serial processes and parallel processes. In serial processes, artifacts are created and brought into a social venue so that others can build on them. During parallel processes, components are produced separately then brought together and combined into one new product. Something new could be jointly created by
multi-agents simultaneously (Fischer et al.,2005). Analogically, a complicated utterance
could be generated by the cooperation of multi-agents obeying a chain rule, a star-like rule or a mixture of both. In the chain rule, one agent initiates a small utterance, then other agents one by one adjust or extend the utterance changed by the previous agent. In the star-like rule, each agent generates its own part of the utterance and all of these would then be combined in various ways.
With the development of artificial languages for creative design, two types of language games are conducted: guessing games between speaker-agents and listener-agents to
evolve grounded language; and generation games between client-agents and designer- agents seeking to generate creative designs in the evolution of compositional languages.
3.3.1 Communication in Guessing Games
speaker listener
2. guessed topic
1. utterance
3
. evaluation
Figure 3.4: Guessing game (afterSaunders and Grace(2008))
Normally, two individuals play the roles of speaker and listener respectively when they
communicate in guessing games (see Fig. 3.4) to evolve a grounded language based on
context and previous simple utterances initially generated via random selection7. They
would develop and share a domain language representing certain topics such as colour and shape by continuing to name what they see, “talk about” what they get and learn from each other. The general process of a guessing game is described as follows.
Firstly, the speaker-agent As selects a topic Ts from the context Ct, which contains a
number of topics such as shapes and colours, and describes the topic Ts with an utter-
ance Us. Then this utterance is sent to the listener-agent Al who parses it to a guessed
topic Tl based on the context Ct. When speaker As receives listener’s guessed topic Tl,
it compares its own topic Ts with the listener’s topic Tl. If the guessed topic is accepted
by the speaker according to its criteria such as the same or in the same category, both
speaker and listener learn (see Equation6) their own representations of the association
between topic (instance) and descriptions, including the extracted meaning and gener-
ated utterance (As: learn(association Ts Us), Al: learn(association Tl Us)). Otherwise,
the listener unlearns (see Equation7) its representation (Al: unlearn(association TlUs))
and learns the association between the speaker’s selected topic with related meaning
and utterance (Al: learn(association Ts Us)).
3.3.2 Communication in Generation Games
Language can be continually developed and applied in generation games played by a
client-agent and several designer-agents (see Fig. 3.5) based on the grounded language
evolved in guessing games. In each generation game, the client-agent Acgenerates a new
utterance Uc by crossing over, mutating or elaborating existing utterances according
7In early stage of guessing games, most utterances are generated randomly to represent selected
topics due to agents’ initial empty memories. Once one or more associations between utterances and topics are memorised by participant-agents in the previous guessing games, they will try to find relevant utterances from these memorised associations in next guessing games. If relevant utterances can not be found, they will still be generated randomly. The mappings of utterances to topics are thus transformed from random connections to relevant associations gradually according to agents’ accumulated experiences during guessing games.
client
designers
2. design
3
. evalu1. briefation
Figure 3.5: Generation game (afterSaunders and Grace(2008))
to the frequency of current utterances or the weights of various associations between
prototypes Pc, meanings Mc and utterances Uc; in other words, based on its memorised
associations and the criteria of evaluating the interestingness of utterances or prototypes.
When the designer-agent Ad receives the client’s requirement, i.e., the new generated
utterance Uc, it generates a new design Ddby parsing the utterance to relevant meanings
Mdwhich could relate to certain prototypes Pdin terms of its own criteria of evaluating
novelty and utility based on previous experience. After all designs, which are likely to be different from each other due to designers’ different experiences, are generated and submitted to the client-agent, the client-agent evaluates each design according to its criteria of creativity related to hedonic functions or certain definitions of specific interesting concepts such as incongruity and exaggeration. If a designer-agent’s work is accepted, both the client-agent and the designer-agent learn the new association between
the design-works and requirement (Ac: learn(association Dd Uc), Ad: learn(association
Dd Uc)). Otherwise, the designer-agent would unlearn its representation of the design
(Ad: unlearn(association Dd Uc)) and try to learn some other designer-agents’ works,
in particular the best design selected by the client-agent (Ad: learn(association Dwinner
Uc)) (Saunders,2011).
The relationship between client-agents and designer-agents, and the mechanics of their communication can be investigated to improve the efficiency of transforming clients’ description language, i.e., requirements, to design concepts. During the repeated gener- ation games, some complex linguistic combinations may emerge to support complicated design briefs encouraging creative meanings.