Evolution of Social Representation in Neural Networks
Evolution of Social Representation in Neural Networks
复制标题
DOI：
10.7551/978-0-262-31709-2-ch061
复制
发表时间：
2013-09
期刊：
影响因子：
--
通讯作者：
S. Arnold;Reiji Suzuki;Takaya Arita
中科院分区：
文献类型：
--
作者：
S. Arnold;Reiji Suzuki;Takaya Arita
本文描述了一种人工生命的方法，以理论的心理（ToM），能够采用心理表征的其他心灵，以了解或预测他人的行为。我们设计了一个模型，在这个模型中，一群神经网络（NN）代理进化出预测能力，基于对过去行为的观察，在新的情况下预测其他人的未来行为。由于代理人的行为是由私人心理状态，不可见的预测代理，这个任务迫使代理超越模仿和重复的配合反应，要求他们获得一定程度的洞察到合作伙伴代理人的内部配置通过观察他们的外部可见的行为。因此，这种学习能力不能用基于奖励或示例的传统学习算法来捕获。我们发现，配备了神经调节机制的神经网络可以进化到在这项任务上表现良好。由此产生的网络表现得好像它们具有一阶ToM的原始形式。
This paper describes an Artificial Life approach to Theory of Mind (ToM), the ability to employ mental representations of other minds in order to understand or anticipate the behaviour of others. We designed a model in which a population of neural network (NN) agents evolve the ability to predict, on basis of observation of past behaviour, others' future behaviour in novel circumstances. As agent behaviour is guided by private mental states, invisible to the predicting agent, this task forces agents to go beyond imitation and repetition of fit responses, requiring them to gain some degree of insight into the partner agent's internal configuration by observation of their externally visible behaviour. As such, this learning ability cannot be captured with conventional learning algorithms based on rewards or examples. We find that NNs equipped with neuromodulation mechanisms can be evolved to perform favourably on this task. The resulting networks are seen to behave as though they have a primitive form of first order ToM.