DOCSLIB.ORG

Multitasking Inhibits Semantic Drift

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Multitasking Inhibits Semantic Drift Multitasking Inhibits Semantic Drift Athul Paul Jacob Mike Lewis Jacob Andreas [email protected] [email protected] [email protected] MIT CSAIL Facebook AI Research MIT CSAIL Abstract Instructor for Task 1 Single-task Executor (build dragons to win) build a When intelligent agents communicate to ac- spearman complish shared goals, how do these goals Standard training executor builds a shape the agents’ language? We study the dragon dynamics of learning in latent language poli- Instructor for Task 1 (build dragons to win) build a Multitask cies (LLPs), in which instructor agents gener- dragon Executor ate natural-language subgoal descriptions and executor agents map these descriptions to low- send a executor explores peasantattack with to with peasant level actions. LLPs can solve challenging exploredragon long-horizon reinforcement learning problems and provide a rich model for studying task- attack with Multitask spearman training oriented language use. But previous work has Instructor for Task 2 executor attacks found that LLP training is prone to seman- (build spearmen to win) with spearman tic drift (use of messages in ways inconsis- (a) (b) (c) (d) tent with their original natural language mean- Figure 1: In latent language policies, instructor agents ings). Here, we demonstrate theoretically and (a) send natural-language commands (b) to executor empirically that multitask training is an effec- agents (c), which execute them in an interactive envi- tive counter to this problem: we prove that ronment (d). Jointly trained instructor–executor pairs multitask training eliminates semantic drift in learn to use messages in ways inconsistent with their a well-studied family of signaling games, and natural language meanings (top shows a real message– show that multitask training of neural LLPs action pair from a model described in Section4). We in a complex strategy game reduces drift and show that multitask training with a population of task- while improving sample efficiency. specific instructors stabilizes message semantics and in some cases improves model performance. 1 Introduction A major goal in the study of artificial and natu- However, they present a number of challenges ral intelligence is to understand how language can for training. As LLPs employ a human-specified scaffold more general problem-solving skills (e.g. space of high-level commands, they must be initial- Spelke, 2017), and how these skills in turn shape ized with human supervision, typically obtained by language itself (e.g. Gibson et al., 2017). In NLP pretraining the executor. On its own, this training paradigm restricts the quality of the learned execu- arXiv:2104.07219v1 [cs.CL] 15 Apr 2021 and machine learning, latent language policies (LLPs; Andreas et al., 2018) provide a standard tor policy to that exhibited in (possibly suboptimal) framework for studying these questions. An LLP human supervision. For tasks like the real-time consists of instructor and executor subpolicies: the strategy game depicted in Fig.1, we would like instructor generates natural language messages (e.g. to study LLPs trained via reinforcement learning high-level commands or subgoals), and the execu- (RL), jointly learning from a downstream reward tor maps these messages to sequences of low-level signal, and optimizing both instructors and execu- actions (Fig.1). LLPs have been used to construct tors for task success rather than fidelity to human interactive agents capable of complex reasoning teachers. (e.g. programming by demonstration) and planning Training LLPs via RL has proven difficult. Past over long horizons (e.g. in strategy games; Hu et al., work has identified two main challenges: primar- 2019). They promise an effective and interpretable ily, the LLP-specific problem of semantic drift, in interface between planning and control. which agents come to deploy messages in ways inconsistent with their original (natural language) goals and communicative needs can facilitate (and meanings (Lewis et al., 2017; Lee et al., 2019); specifically stabilize) learning of communication secondarily, the general problem of sample inef- strategies. ficiency in RL algorithms (Kakade et al., 2003; Brunskill and Li, 2013). Model-free deep RL is par- 2 Background and Related Work ticularly notorious for requiring enormous amounts of interaction with the environment (Munos et al., Deep reinforcement learning (DRL) has recently 2016; Mnih et al., 2013b). For LLPs to meet their made impressive progress on many challenging do- promise as flexible, controllable, and understand- mains such as games (Mnih et al., 2013a; Silver able tools for deep learning, better approaches are et al., 2016), locomotion (Schulman et al., 2015) needed to limit semantic drift and perhaps improve and dexterous manipulation tasks (Gu et al., 2016; sample efficiency. Rajeswaran et al., 2017). However, even state-of- While semantic change is a constant and well- the-art approaches to reinforcement struggle with documented feature of human languages (McMa- tasks involving complex goals, sparse rewards, and hon and April, 1994), (human) word meanings are long time horizons. A variety of models and algo- on the whole remarkably stable relative to the rate rithms for hierarchical reinforcement learning have of change in the tasks for which words are deployed been proposed to address this challenge (Dayan (Karjus et al., 2020). In particular, disappearance and Hinton, 1993; Dietterich, 2000; Richard et al., of lexical items is mitigated by increased popu- 1999; Bacon et al., 2017) via supervised or unsuper- lation size (Bromham et al., 2015) and increased vised training of a fixed, discrete set of sub-policies. frequency of use (Pagel et al., 2007). Drawing Language can express arbitrary goals, and has on these facts about stabilizing factors in human compositional structure that allows generalization language, we hypothesize that training of machine across commands. Building on this intuition, sev- learning models with latent language variables can eral recent papers have explored hierarchical RL be made more robust by incorporating a population in which natural language is used to parameterize of instructors with diverse communicative needs the space of high-level actions (Oh et al., 2017; that exercise different parts of the lexicon. Andreas et al., 2017; Shu et al., 2018; Jiang et al., We describe a multitask LLP training scheme in 2019; Hu et al., 2019). While there are minor imple- which task-specific instructors communicate with mentation differences between all these approaches, a shared executor. We show that complex long- we will refer to them collectively as latent lan- horizon LLPs can be effectively tuned via joint guage policies (LLPs). Like other hierarchical reinforcement learning of instructors and executors agents, an LLP consists of a pair of subpolicies: an using multitask training: instructor I(m j o) and an executor E(a j m; o). An LLP takes actions by first sampling a string- • Section3 presents a formal analysis of LLP valued message m ∼ I from the instructor, and training as an iterated Lewis signalling game then an action a ∼ E from the executor. For these (Lewis, 1969). By modeling learning in this messages to correspond to natural language, rather game as a dynamical system, we completely than arbitrary strings, policies need some source characterize a class of simple policies that are of information about what human language users subject to semantic drift. We show that a par- think they mean. This is typically accomplished ticular multitask training scheme eliminates by pretraining executors via human demonstrations the set of initializations that undergo semantic or reinforcement learning; here we focus on the drift. ingredients of effective joint RL of instructors and • Section4 evaluates the empirical effectiveness executors. of multitask learning in a real-time strategy Reinforcement learning has been widely used to game featuring rich language, complex com- improve supervised language generation policies, plex dynamics, and LLPs implemented with particularly for dialogue (Li et al., 2016; Lewis deep neural networks. Again, we show that et al., 2017), translation (Ranzato et al., 2015; Wu multitask training reduces semantic drift (and et al., 2016) and summarization (Stiennon et al., improves sample efficiency) of LLPs in multi- 2020). Here, we instead focus on models where ple game variants. language is a latent variable as part of a hierarchical Together, these results show that diverse shared policy for a non-linguistic task. Instructor I m o Executor E a m ( ∣ ) ( ∣ ) a a a R 1 2 R a1 2 push red a ′ o m1 = 1 1 1 0 0 1 o 1 o1 push blue m2 = a2 0 1 1 0 o 2 o2 o2 Figure 2: A signaling game with two possible observations, two possible messages, and two possible actions. The instructor observes either a triangle or a square, then sends a message to a executor, who pushes either the red or blue buttons. The players’ reward depends on the observation and the action but not the message. Two possible reward functions, R and R0, are shown at right. As noted in Section1, an observed shortcoming ing improves the faithfulness and adaptability of of reinforcement learning in all these settings is its learned language understanding models, even when susceptibility to semantic drift. In the literature optimizing for a downstream reward. on human language change (Blank, 1999), seman- tic drift refers to a variety of phenomena, includ- 3 Multitask Communication in Theory: ing specific terms becoming more general, general Lewis Signaling Games terms becoming specific, and parts coming to refer We begin our analysis with the simple signaling to wholes. In machine learning, it refers broadly to game depicted in Fig.2. In this game, one agent the use of messages inconsistent with their natural receives an observation, then sends a message to language meanings in language-generation policies another agent, which then performs an action. Sig- (Lazaridou et al., 2020). naling games like this one are widely studied in Lee et al.(2019) mitigate semantic drift in pivot- NLP as models of reference resolution and lan- based machine translation by using visual ground- guage generation (Frank and Goodman, 2012).
Load more

Recommended publications
  • CONTENT2 2.Pages
    DRAFT PROPOSITIONAL CONTENT in SIGNALS (for Special Issue Studies in the History and Philosophy of Science C ) Brian Skyrms and Jeffrey A. Barrett Introduction We all think that humans are animals, that human language is a sophisticated form of animal signaling, and that it arises spontaneously due to natural processes. From a naturalistic perspective, what is fundamental is what is common to signaling throughout the biological world -- the transfer of information. As Fred Dretske put it in 1981, "In the beginning was information, the word came later." There is a practice of signaling with information transfer that settles into some sort of a pattern, and eventually what we call meaning or propositional content crystallizes out. The place to start is to study the evolution, biological or cultural, of information transfer in its most simple and tractable forms. But once this is done, naturalists need first to move to evolution of more complex forms of information transfer. And then to an account of the crystallization process that gives us meaning. There often two kinds of information in the same signal: (1) information about the state of the world that the signaler observes (2) information about the act that the receiver will perform on receiving the signal. [See Millikan (1984), Harms (2004), Skyrms (2010)]. The meaning that crystalizes out about the states, we will single out here as propositional meaning. The simplest account that comes to mind will not quite do, as a general account of this kind of meaning, although it may be close to the mark in especially favorable situations.
    [Show full text]
  • Information Games and Robust Trading Mechanisms
    Information Games and Robust Trading Mechanisms Gabriel Carroll, Stanford University [email protected] October 3, 2017 Abstract Agents about to engage in economic transactions may take costly actions to influence their own or others’ information: costly signaling, information acquisition, hard evidence disclosure, and so forth. We study the problem of optimally designing a mechanism to be robust to all such activities, here termed information games. The designer cares about welfare, and explicitly takes the costs incurred in information games into account. We adopt a simple bilateral trade model as a case study. Any trading mechanism is evaluated by the expected welfare, net of information game costs, that it guarantees in the worst case across all possible games. Dominant- strategy mechanisms are natural candidates for the optimum, since there is never any incentive to manipulate information. We find that for some parameter values, a dominant-strategy mechanism is indeed optimal; for others, the optimum is a non- dominant-strategy mechanism, in which one party chooses which of two trading prices to offer. Thanks to (in random order) Parag Pathak, Daron Acemoglu, Rohit Lamba, Laura Doval, Mohammad Akbarpour, Alex Wolitzky, Fuhito Kojima, Philipp Strack, Iv´an Werning, Takuro Yamashita, Juuso Toikka, and Glenn Ellison, as well as seminar participants at Michigan, Yale, Northwestern, CUHK, and NUS, for helpful comments and discussions. Dan Walton provided valuable research assistance. 1 Introduction A large portion of theoretical work in mechanism design, especially in the world of market design applications, has focused on dominant-strategy (or strategyproof ) mechanisms — 1 those in which each agent is asked for her preferences, and it is always in her best interest to report them truthfully, no matter what other agents are doing.
    [Show full text]
  • Arxiv:1612.07182V2 [Cs.CL] 5 Mar 2017 in a World Populated by Other Agents
    Under review as a conference paper at ICLR 2017 MULTI-AGENT COOPERATION AND THE EMERGENCE OF (NATURAL)LANGUAGE Angeliki Lazaridou1∗, Alexander Peysakhovich2, Marco Baroni2;3 1Google DeepMind, 2Facebook AI Research, 3University of Trento [email protected], falexpeys,[email protected] ABSTRACT The current mainstream approach to train natural language systems is to expose them to large amounts of text. This passive learning is problematic if we are in- terested in developing interactive machines, such as conversational agents. We propose a framework for language learning that relies on multi-agent communi- cation. We study this learning in the context of referential games. In these games, a sender and a receiver see a pair of images. The sender is told one of them is the target and is allowed to send a message from a fixed, arbitary vocabulary to the receiver. The receiver must rely on this message to identify the target. Thus, the agents develop their own language interactively out of the need to communi- cate. We show that two networks with simple configurations are able to learn to coordinate in the referential game. We further explore how to make changes to the game environment to cause the “word meanings” induced in the game to better re- flect intuitive semantic properties of the images. In addition, we present a simple strategy for grounding the agents’ code into natural language. Both of these are necessary steps towards developing machines that are able to communicate with humans productively. 1 INTRODUCTION I tried to break it to him gently [...] the only way to learn an unknown language is to interact with a native speaker [...] asking questions, holding a conversation, that sort of thing [...] If you want to learn the aliens’ language, someone [...] will have to talk with an alien.
    [Show full text]
  • Game Theory and Scalar Implicatures
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by UCL Discovery Game Theory and Scalar Implicatures Daniel Rothschild∗ September 4, 2013 1 Introduction Much of philosophy of language and linguistics is concerned with showing what is special about language. One of Grice’s (1967/1989) contributions, against this tendency, was to treat speech as a form of rational activity, subject to the same sorts of norms and expectations that apply to all such activity. This gen- eral perspective has proved very fruitful in pragmatics. However, it is rarely explicitly asked whether a particular pragmatic phenomenon should be under- stood entirely in terms of rational agency or whether positing special linguistic principles is necessary. This paper is concerned with evaluating the degree to which a species of simple pragmatic inferences, scalar implicatures, should be viewed as a form of rational inference. A rigorous answer to this last ques- tion requires using the theoretical resources of game theory. I show that weak- dominance reasoning, a standard form of game-theoretic reasoning, allows us to cash out the derivation of simple scalar implicatures as a form of rational infer- ence. I argue that this account of the scalar implicatures is more principled and robust than other explanations in the game theory and pragmatics literature. However, we can still see that deriving Gricean implicatures on the basis of decision and game-theoretic tools is not nearly as straightforward as we might hope, and the ultimate tools needed are disquietingly powerful. ∗ This paper is a distant descendent of my first, tenative foray into these issues, “Grice, Ratio- nality, and Utterance Choice”.
    [Show full text]
  • Responsiveness and Robustness in the David Lewis Signaling Game
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by The Australian National University Responsiveness and Robustness in the David Lewis Signaling Game Carl Brusse and Justin Bruner*y We consider modifications to the standard David Lewis signaling game and relax a num- ber of unrealistic implicit assumptions that are often built into the framework. In par- ticular, we motivate and explore various asymmetries that exist between the sender and receiver roles. We find that endowing receivers with a more realistic set of responses sig- nificantly decreases the likelihood of signaling, while allowing for unequal selection pres- sure often has the opposite effect. We argue that the results of this article can also help make sense of a well-known evolutionary puzzle regarding the absence of an evolution- ary arms race between sender and receiver in conflict-of-interest signaling games. 1. Signaling Games and Evolution. Common interest signaling games were introduced by David Lewis (1969) as a game theoretic framework that treats communicative conventions as solutions to coordination problems. In recent years, this has informed a growing body of work on the evolution of communication, incorporating signaling games into an evolutionary game theoretic approach to modeling the evolution of communication and coop- eration in humans (Skyrms 1996, 2010). As the basis for such work, Lewis signaling games are attractive in their intuitive simplicity and clear outcomes. They are coordination games be- tween world-observing senders and action-making receivers using costless signals, in contrast to games where interests may differ and where costly sig- nals are typically invoked.
    [Show full text]
  • Signals and the Structure of Societies Dissertation Roland Mühlenbernd
    Signals and the Structure of Societies University of T¨ubingen Department for General Linguistics Dissertation in Linguistics by Roland M¨uhlenbernd Promotor: Professor Dr. Gerhard J¨ager T¨ubingen,2013 . - Contents Contentsi Acknowledgementsv Introduction1 1 Signaling Games9 1.1 Definition of a Signaling Game................. 12 1.2 Strategies, Equilibria & Signaling Systems.......... 14 1.3 The Horn Game......................... 17 1.4 Repeated Games ........................ 21 1.5 Conclusion............................ 23 2 Update Dynamics 25 2.1 Evolutionary Dynamics..................... 27 2.1.1 Symmetric and Asymmetric Static Signaling Games . 28 2.1.2 Evolutionary Stability ................. 30 2.1.3 Replicator Dynamics.................. 31 2.1.4 The Lewis Game in Evolution............. 33 2.1.5 The Horn Game in Evolution ............. 36 2.1.6 The Plausible Symmetric Horn Game......... 43 2.2 Imitation Dynamics....................... 48 2.2.1 Imitate the Best..................... 48 2.2.2 Conditional Imitation.................. 50 2.2.3 Comparing Replicator Dynamics and Imitation . 51 2.2.4 Integrating Network Structure............. 53 2.2.5 Imitation on a Grid: Some Basic Results . 55 2.3 Learning Dynamics....................... 57 i ii CONTENTS 2.3.1 Languages, Learners and Stability........... 59 2.3.2 Reinforcement Learning ................ 63 2.3.3 Belief Learning ..................... 65 2.3.4 Memory Size and Forgetting.............. 68 2.3.5 Learning in Populations: Some Basic Results . 69 2.4 Conclusion............................ 72 3 Social Networks 75 3.1 Network Properties....................... 76 3.1.1 Node Properties..................... 76 3.1.2 Structural Properties.................. 79 3.2 Network Types ......................... 83 3.2.1 Regular Networks.................... 83 3.2.2 Random Networks ................... 85 3.2.3 Small-World Networks................
    [Show full text]
  • [7] Skyrms (2010) Signals.Pdf
    Signals This page intentionally left blank Signals Evolution, Learning, & Information Brian Skyrms 1 3 Great Clarendon Street, Oxford OX26DP Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide in Oxford New York Auckland Cape Town Dar es Salaam Hong Kong Karachi Kuala Lumpur Madrid Melbourne Mexico City Nairobi New Delhi Shanghai Taipei Toronto With offices in Argentina Austria Brazil Chile Czech Republic France Greece Guatemala Hungary Italy Japan Poland Portugal Singapore South Korea Switzerland Thailand Turkey Ukraine Vietnam Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries Published in the United States by Oxford University Press Inc., New York # Brian Skyrms 2010 The moral rights of the author have been asserted Database right Oxford University Press (maker) First published 2010 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this book in any other binding or cover and you must impose the same condition on
    [Show full text]
  • Modelling Religious Signalling
    1 MODELLING RELIGIOUS SIGNALLING A thesis submitted for the degree of Doctor of Philosophy at the Australian National University Carl Joseph Brusse February 2019 © Copyright by Carl Joseph Brusse, 2019 All Rights Reserved 2 3 This thesis is solely the work of its author. No part of it has been submitted for any degree, or is currently being submitted for any other degree. To the best of my knowledge, any help received in preparing this thesis, and all sources used, have been duly acknowledged. 4 5 for Caroline, Alida, and Isobel 6 7 ACKNOWLEDGEMENTS I owe a great debt to my supervisory panel members past and present: Kim Sterelny, Justin Bruner, Rachael Brown, and Ben Fraser, for their deeply insightful commentary and gentle corrective during the various phases of this project. In particular, Kim very much provided both the inspiration for this project and the impetus for venturing to return to philosophy for this PhD candidature. I would not have entered into this without his encouragement, and I certainly would not have made it out the other side without his support and mentorship. Thank you, Kim, for your judgment and generosity, and for your astute, necessity-sensitive application of patience (or lack thereof) when required. Justin too deserves special mention for patiently schooling me in evolutionary game theory and signalling games (despite my sometimes bizarrely misguided intuitions), and in suggesting and contributing to my first simulation projects and our joint paper at the 2016 PSA, published in Philosophy of Science. Being at the School of Philosophy at the ANU, the number of faculty and grad students (local and visiting) who have taken the time to listen and offer valuable insights would be too long to record here.
    [Show full text]
  • Games in the Philosophy of Biology
    Games in the Philosophy of Biology Cailin O'Connor July 20, 2019 2 Chapter 1 Introduction A male peacock mantis shrimp resides happily in his burrow. He has recently molted, leaving his shell soft and vulnerable. As he waits, another male wanders onto his territory and approaches. The new male would like a burrow of his own. Both possess raptorial claws powerful enough to smash through the glass of an aquarium. If they fight for the territory, the temporarily squishy burrow owner will be seriously hurt. Neither, though, can directly observe whether the other has molted recently. Both mantis shrimp raise their appendages to display brightly colored `meral spots', intended to signal their aggression and strength. The intruder is impressed, and backs off to seek territory elsewhere. Alex is looking to hire a new data analyst for his company. Annaleigh wants the job, and so is keen to impress Alex. She includes every accomplishment she can think of on her resume. Since she went to a prestigious college, she makes sure her educational background is front and center where Alex will see it first. A group of vampire bats need to eat nightly to maintain their strength. This is not always easy, however. Sometimes a bat will hunt all night but fail to find a meal. For this reason, most bats in the group have established relationships for reciprocal food sharing. Bats who manage to feed will regurgitate blood for partners who did not. Mitzi needs a kidney transplant, but is too old to go on the normal donor list.
    [Show full text]
  • Common Interest and Signaling Games: a Dynamic Analysis
    Common Interest and Signaling Games: A Dynamic Analysis Manolo Martínez and Peter Godfrey-Smith* We present a dynamic model of the evolution of communication in a Lewis signaling game while systematically varying the degree of common interest between sender and receiver. We show that the level of common interest between sender and receiver is strongly predictive of the amount of information transferred between them. We also discuss a set of rare but interesting cases in which common interest is almost entirely absent, yet substantial information transfer persists in a “cheap talk” regime, and offer a diagnosis of how this may arise. 1. Introduction. An important recent development in the naturalistic study of communication is the systematic investigation of simple, computation- ally tractable models. Such models are severely idealized in many respects, but there are invaluable gains in the explicitness and rigor of the results ob- tained by working on them, alongside the familiar approach of engaging in informal discussion of more realistic examples. Some areas of concern to philosophers that this research program has already shed light on are the difference between assertions (indicatives) and directives (imperatives) (Hut- tegger 2007; Zollman 2011), signaling in social dilemmas (Wagner 2015), de- ception (Zollman, Bergstrom, and Huttegger 2013; Martínez 2015), and vague- ness (O’Connor 2014). Formulating the problem of communication in a way that makes it ame- nable to a rigorous treatment of this kind is in itself a major philosophical contribution. Most of the work cited above is based on Lewis’s (1969/2002) model of signaling. In this model a sender (or, for Lewis, ‘communicator’) sends messages to a receiver (or, for Lewis, an ‘audience’), and both parties Received May 2015; revised July 2015.
    [Show full text]
  • Some Dynamics of Signaling Games
    Some dynamics of signaling games Simon Hutteggera, Brian Skyrmsa,1, Pierre Tarrèsb, and Elliott Wagnerc aDepartment of Logic and Philosophy of Science, University of California, Irvine, CA 92697; bDepartment of Mathematics, University of Toulouse, 31000 Toulouse, France; and cDepartment of Philosophy, Kansas State University, Manhattan, KS 66506 Edited by John C. Avise, University of California, Irvine, CA, and approved April 4, 2014 (received for review February 7, 2014) Information transfer is a basic feature of life that includes signal- communication. Costly signaling games are studied in economics, ing within and between organisms. Owing to its interactive na- starting with the Spence game (6), and in biology (e.g., ref. 4). The ture, signaling can be investigated by using game theory. Game Spence game is a model of job market signaling. In a job market, theoretic models of signaling have a long tradition in biology, employers would like to hire qualified job candidates, but the level economics, and philosophy. For a long time the analyses of these of a job candidate’s qualification is not directly observable. In- games has mostly relied on using static equilibrium concepts such stead, a job candidate can send a signal about her qualification to as Pareto optimal Nash equilibria or evolutionarily stable strate- the employer. However, if signals are costless, then job candidates gies. More recently signaling games of various types have been will choose to signal that they are highly qualified regardless of investigated with the help of game dynamics, which includes their qualification. If a signal is more costly for low-quality can- dynamical models of evolution and individual learning.
    [Show full text]
  • Evolutionary Explanations of Simple Communication Signalling Games & Their Models
    Journal for General Philosophy of Science. Pre-print. Please cite published version. The final publication is available doi.org/10.1007/s10838-019-09481-7 Evolutionary Explanations of Simple Communication Signalling Games & Their Models Travis LaCroix Received: 8 March 2019 / Accepted: 8 July 2019 Abstract This paper applies the theoretical criteria laid out by D'Arms et al. (1998) to various aspects of evolutionary models of signalling. The question that D'Arms et al. seek to answer can be formulated as follows: Are the models that we use to explain the phenomena in question conceptually adequate? The conceptual adequacy question relates the formal aspects of the model to those aspects of the natural world that the model is supposed to capture. More- over, this paper extends the analysis of D'Arms et al. by asking the following additional question: Are the models that we use sufficient to explain the phe- nomena in question? The sufficiency question asks what formal resources are minimally required for the model to get the right results most of the time. Keywords Signalling Games · Evolutionary Game Theory · Evolutionary Models · Robustness · Modelling 1 Introduction To communicate meaningfully, members of a linguistic population must co- operate|i.e., they must agree upon some convention: one individual should use a signal to mean one thing if (most) other individuals use that signal in the same way. Conventional language-use in a population of speakers can be modelled using game-theoretic tools; this idea gives rise to the signalling- game
    [Show full text]