Daily experiences influence our learning and change the way we think and act. Sometimes we are not even aware that we are learning from our surroundings, which is a very informal way of perceiving things. On the other hand, we can also learn in a formal way from a a structured classroom environment. Learning is not limited to acquiring knowledge or facts; we also learn skills and attitudes. This can happen in different ways. We learn new ideas and concepts from a lecture or a discussion, whereas skills must be acquired via continuous practice and receiving simultaneous feedback from an instructor. In a planned environment, learning is reinforced by teachers who expect students to memorize the content and later reward them for it. In contrast, researchers or scientists learn by investigating things themselves, over time. But in any form of learning, motivations and rewards play a very important role, as people derive satisfaction from the feeling of competence. In the case of learning a new skill, people can be strongly motivated by the incentives they are given, which might lead them to acquire new knowledge which they can use in future.
Teachers are entrusted with the job of determining what the student will learn. They are not only instrumental to the students’ learning, but also make sure that they have learnt the subject properly. Teaching must be planned very carefully, taking the learning styles and the background of the students into account. Teachers also need to assess students often to determine how well they are progressing and simultaneously attend to their weaknesses. Hence teaching and learning are constructed over a series of intrinsic and extrinsic social interactions which influence the cognitive models of both the teacher and the learner. The learning can be improved if the facilitator can teach each individual, possessing some understanding of the subtleties of the student’s mind, behavior and learning style, and regulating the motivational strategies accordingly (Prozesky, 2000).
In this research, the humanoid robot Baxter and a computer system (depending on the experiment) motivate an individual extrinsically during the learning process using several positive reinforcers. During the interactions, the robot or the system initially demonstrates the task to the participant, and then provides learned reinforcers to make sure the skill is transferring properly, using Simpson’s psychometric model (Simpson, 1972), and concurrently learns about their cognitive models. The expert uses a reinforcement learning strategy to understand the effect of its reinforcement presentation on its human subjects, attempting to increase their performance over time (Roy et al., 2019b, a, 2018b). To identify the success of MRL-guided skill transfer, we divided the subject population into three groups where participants get no reinforcement, random reinforcement, or individually-tailored learned reinforcement (MRL) respectively. We compared the number of mistakes in each group, because we expect MRL to be more effective for subjects who are performing somewhat poorly. We discovered that participants in the learned group were more likely to perform well and committed comparatively fewer mistakes with respect to the other experimental conditions ( for Baxter,
for the computer-based Tetris skill learning). We also determined information gain over time and how a machine’s regret strongly correlates with the probability that a test subject makes more versus fewer mistakes. In addition, we produced confusion matrices demonstrating the effectiveness of MRL in the experiments using 5-point Likert scale data.
2. Related work
Many contemporary researchers are working on identifying appropriate reward channels in human-robot collaborative frameworks to maximize performance. The following section briefly discusses this work along with the influence of positive reinforcers on motivating humans in skill transfer procedures.
2.1. Reinforcement learning techniques to identify better reward channels
Rewards play a crucial role in both identifying and shaping a person’s behavior. They not only tell us about a person’s personality, but provide an influence channel when used effectively. Hence, recently many scientists are interested in researching appropriate reward channels that might increase task performance. Lopes (Lopes et al., 2013) upgraded Multi-Arm Bandit techniques using different motivational resources to maximize skills and learning activities. He also researched the recovery of reward functions from expert demonstrated policies (Lopes et al., 2009)
to ensure active learning. The modification of traditional reinforcement learning algorithms using reward shaping produced important insights into how skill and accuracy can be improved for a particular task. Cooperative inverse reinforcement learning (CIRL) uses a human reward functionthat maps world states, joint actions, and reward parameters to real numbers to establish useful human robot collaboration, where the robot is unaware of the initial reward. CIRL can be used in various platforms like active teaching, active learning, and communicative actions that are more effective in achieving value alignment (Hadfield-Menell et al., 2016). These researchers are also using multi-arm bandit techniques (MAB) to address the problems with assistive agents who can help human participants to select appropriate channels to maximize the cumulative reward (Chan et al., 2019)
. Here the human does not know the reward function but can learn it through several interactions, whereas the robot only observes the human interactions and not the reward associated with it. Tabrez et al. used their Reward Augmentation and Repair through Explanation (RARE) framework for estimating task understanding where the autonomous agent detects potential causes of system failures and uses human-interpretable feedback for model correction(Tabrez et al., 2019). Nikolaidas et al.(Nikolaidis and Shah, 2013) described a human-robot cross-training framework using reinforcement learning techniques where humans and robots switch roles to improve the overall performance. Li et al. used MRL in automatic poetry generation using two models (local and global) which have some predefined criteria as rewards, and they learn from each other to pursue higher scores (Yi et al., 2018). Griffith (Griffith et al., 2013) discussed novel policy shaping algorithms and how motivations and reward signals can be used as a channel to impact human-robot partnerships in an HRI setting, simultaneously improving the future learning process of both humans and robots. Knox et al. (Knox and Stone, 2009; Bradley Knox and Stone, 2008) designed a novel framework named TAMER which allows a human to train a learning agent to perform a complex tasks over continuous interaction. In our previous papers (Roy et al., 2018b) we have also discussed how the robot updates its own cognitive model with each human interaction, improving the overall task performance through exploration-exploitation strategies. In this work we are not only extending the goal of our previous work, but also updating the MRL technique for better results.
2.2. Robots learning to teach
Many scientists have started exploring this new area of robotics, where along with robot teaching we can gain other useful information about robot and human behaviours from their interaction. Spaulding (Spaulding et al., 2018) introduced an integrated system for autonomously analyzing and assessing children’s speech and pronunciation in the context of an interactive word game between a social robot and a child. This approach used Gaussian Process Regression (GPR), augmented with an active Learning protocol that informed the robot’s behavior. Scassellati (Ramachandran and Scassellati, 2015; Litoiu and Scassellati, 2015) and Park (Park et al., 2017) have presented feedback-based human-robot interaction, demonstrating that if humans are guided by a robot at an interpersonal level, it increases the robot’s perceived social reliability and makes humans more eager to interact with it. A robot learning from human feedback tends develop a mental model (Lee et al., 2005; Scassellati, 2001) of its own which positively influences human cognition. Fasola et al. (Fasola and Matarić, 2013) used socially assistive robots (SAR) to train elderly humans in physical fitness by motivating them. Yin et al. (Yin et al., 2015) described intelligent robot systems acquiring human-like writing style and then exploiting it to teach children. Fan et al. (Fan et al., 2018)
used neural network models to evaluate teaching strategies when one intelligent system is trying to teach another. Leite et al.(Leite et al., 2012a) used robots to socially support children in a game scenario. The robot not only increased the performance accuracy of the human learner, but also connected with them emotionally and provided social assistance throughout their learning process. This social support helped the children to build their self esteem and encouraged them to perform better. Cakmak (Cakmak et al., 2009) demonstrated how social learning strategies vary with the particular environment when robots are allowed to explore and learn from their surroundings.
In this paper, along with the effectiveness of MRL, we are also concerned with the idea of robots learning to be good teachers. We use a robot’s own predicted regret and confusion matrices to evaluate its own cognitive model.
2.3. Empathy and positive reinforcers
Empathy is based in the social-cognitive and behavioral ability to vicariously experience another person or animal’s affect, and is critical in the social interactions of humans and some animals (Keskin, 2014; Lockwood, 2016). Empathy plays a vital role in social interaction in all stages of human life and many contemporary researchers are working on empathetic robots that are designed to respond to human behavior and emotion with appropriate social cues. Empathy and adaptation may not be enough, however, since social responses are only one component of effective human-robot interactions. Instead, robot interactions that facilitate mutual learning with the human counterpart may prove more effective in a teaching environment due to the ability to learn, adapt, and create reinforcement feedback tailored to the individual. Hence the ability to empathize has also been found to be a critical characteristic of effective teachers. In one study, teachers demonstrating more empathy were able to adapt the structure, behavior, and manifestation of empathy based on the group or individual and provide more effective teaching strategies (Lazarescu, 2013). In our research, we used several positive reinforcers as reward channels to interact with the human participant.
In order to effectively study human-robot interactions and learning, scientists have incorporated other socially-inspired tools in addition to empathy. For example, auditory and visual cues are important in learning exchanges between humans and robots, especially when learning through demonstration (Koenig et al., 2010). Furthermore, modeling demonstration learning using robots and humans has shown to be effective and the closer the demonstration technique was to typical social learning, the more rapport the participant felt with the robot and the more he or she learned (Sauppé and Mutlu, 2015). Humans also demonstrate a need to share intentions with their social partners, and in order to mimic this with robots, the robot partner needs to mimic the social skills necessary to interact with humans and demonstrate shared intention (Dominey and Warneken, 2011). The robot, in this case, demonstrated the ability to learn a goal and intentional actions linked to the goal through cooperative learning (Dominey and Warneken, 2011). In these cases, behavioral interactions and social acceptance are critical components to the human-robot interaction.
It is possible for humans to respond to perceived empathy from robot and computer interactions. Research shows that individuals perceive empathy through digital devices and computer-mediated interactions (Powell and Roberts, 2016), and additional studies are developing robot-human interactions that more closely mimic human-human interactions using touch and visual interactions (Salter et al., 2006). Furthermore, empathy increases rapport between humans and robots, which is important for user comfort (Leite et al., 2012b). This suggests that, while empathy is important for contextual comfort, it may not be the only component of a learning environment and does not indicate a human response for the robot. While scientists may have developed robots to mimic empathy that can be detected by participants, humans have yet to respond with equal attachment or empathy towards robots (Konok, 2018). In other words, while adaptive empathetic robots may build some rapport with humans, the communication is only from the robot to the human; the robot is not necessarily responding in ways that may be necessary for human learning. A few researchers have explored various areas where positive reinforcement from robots had a large impact on children. Boccanfuso et al. (Boccanfuso et al., 2016) investigated the difference in responses between children with or without autism with an emotion-stimulating robot using positive reinforcement in an interactive environment. Nunez et al. (Nunez et al., 2015) described the use of positive reinforcers to overcome the underlying challenges in motivating a child to continue learning and to share the experience with others. Kim (Kim et al., 2013) addressed the unique positive effects and advantages a robot can have on autistic children, exploring areas where robots play an important role in the lives of specific individuals. We wish to investigate how a robot can develop an understanding of the underlying motivations and cognitive traits of individual people, so that it can shape its teaching strategies appropriately and enhance the learning process.
2.4. Positive reinforcers in MRL
Mutual feedback between robot and human has become increasingly important in human-robot interactions. Interactivism (Bickhard, 2009) and process-oriented robots have been challenging in the past since there is a necessary balance between environmental stimuli and feedback and the adaptation of software and processes that can adapt and change with them (Stojanov et al., 2006). Robots using socially-inspired reinforcement including verbal and behavioral feedback have only shown modest results, and studies have suggested that a more targeted approach tailored to the individual would be better suited for future robot-human interactions (Ferreira and Lefèvre, 2015).
To better facilitate natural social interactions and engage with the learning environments of humans, robots need to adapt and respond appropriately to each individual. Positive reinforcement increases learning in all animals and promotes voluntary behaviors of animals, but the reinforcement tools need to be species-specific and based on individual preferences and experiences. In this sense, if robot-human interactions are to use socially-derived reinforcements as teaching tools, researchers need to take into account not just human social interactions, but individual differences as well. This means that the robots need to be programed with an understanding of individual-specific approaches to interactions based on principles of learning and be able to adapt and respond in ways that are tailored based on the individual’s unique responses.
The scientists in this study have developed a novel approach using mutual reinforcement learning where both the robot and human act as individual empathizers who can act as reinforcement learning agents to achieve a particular task. Thus in this paper the humanoid robot Baxter and a personal computer not only adapts or empathizes with its human participant but also takes a step forward to encourage them and achieve their goal.
3. Technical description
Mutual reinforcement learning (MRL) deals with the scenario where both humans and autonomous agents act as reinforcement learners for each other, identifying the path to achieve maximum reward. In this instance, the robot initially acts as an expert and its human counterpart as a novice. In MRL, one agent’s action is mapped as a reward to another. Here the agents, as they are unaware of each other’s incoming actions, discover the appropriate reward channel over continuous communications with one another. The autonomous agent acting as an expert learns about the appropriate reward channel through an exploration-exploitation tradeoff (Audibert et al., 2009). The action of the novice agent (judged in terms of the agent’s performance) not only affects the immediate rewards but also the expert’s next action. The expert does not immediately jump to a conclusion about the decision to be made, but rather invests more time and effort in accumulating further information, with the hope that a broader perspective will lead to a better decision in the future. On the other hand, the humans interpret the actions of the robot (or computer) agent as a reward, which influences their performance in learning the task. In this paper, MRL is implemented in a skill transfer scenario, where the autonomous agent is trying to teach a human some complex task, while updating its own mental model (Sutton and Barto, 2018) at the same time.
MRL is a tuple where is a set of states; and are sets of actions; is the set of state transition probabilities upon taking action or in state , and and or and are the reward functions. Since, in MRL, the action of an agent is the reward to another and vice versa, the tuple can be simplified as follows: Novice=, Expert= where if the novice executes action , reward is received by the expert. This helps the expert to execute action using an exploration/exploitation strategy, which at the same time acts as a reward to the participant. If the action is successful, then the robot realizes that the participant is fonder of reward , which acts at the same time as reward for the robot to understand its own performance or action . Here the reward for the novice is selected by an exploration-exploitation tradeoff where are all different kinds of reinforcers mentioned in Section 4. In the case of MRL, we have a verbal, hint, gesture and simple feedback for the robot whereas there are seven different reinforcers in case of Tetris. Therefore the expected rewards in both the cases in a state action pair can be written as a two-argument function and .
The above equations imply that whenever a novice makes a mistake at time , the robot takes an action in that state to positively reinforce the participant, who on the other hand takes an action and rectifies the mistake moving to the next state. Here the state is the pattern-making task in the case of Baxter, while in the case of Tetris, the players are asked to restart the game rectifying the mistake. Hence the reinforcement learning agents give rise to a sequence or trajectory that looks like the following if the novice keeps on making a mistake: . The above sequence denotes the condition if a participant keeps on making a mistake at one point. The sequence will stop with the correct action of the novice learner. However, the robot keeps on evaluating the other sections and if a mistake occurs again, the same behavior is repeated. Therefore the state transition probability :
Robotic mutual reinforcement is based on psychological principles of social reinforcement and inclusion and is intended to improve skill transfer by adapting to the reward value systems of an individual. In order to effectively teach a skill, the instructor relies on the principles of learning theory and basic operant conditioning and positive reinforcement. Reinforcement, whether through the addition of a reward (positive reinforcement) or the removal of something aversive (negative reinforcement) refers to techniques used by trainers and instructors whose goal is to increase the likelihood of the behavior being repeated. Behavioral psychologists rely on the principles of positive reinforcement as the primary means through which to teach and shape behaviors in both humans and animals. This type of shaping is considered ideal since the individual participant or subject is rewarded for the correct behavior and associates a specific behavior with a specific reward. The subject is therefore more likely to repeat the desired behavior in the future (as opposed to negative reinforcement when the subject is trialing behaviors to avoid an aversive stimulus). In the case of mutual reinforcement, the robot engages in activities and behaviors that positively reinforce the correct behavior of the participant. Learning theory dictates, however, that the value of rewards in positive reinforcement are subjective in nature and highly dependent upon the individual. This means that a reward () may be of high value to one individual and of no value to another. For adequate learning to occur, the reward must therefore be tailored to each individual. In the case of mutual reinforcement, the robot is equipped with a range of rewards that might be of value to the subject. Given that all the subjects are human, the programmed rewards are tailored to humans and were designed based on the species, culture, and potential individual differences of the target population. Humans are social and, in performance environments, are highly responsive to social inclusion and exclusion (Cheung and Gardner, 2015), making social signals of high value to almost every human. The social rewards are designed to mimic culturally-appropriate interactions to which the participants are accustomed, which promotes comfort and relaxation and builds rapport. The ability to adapt and read the human through behavioral feedback establishes a baseline of communication and language that is novel and unique between the robot and the human, effectively mimicking normal social relationships and social learning paradigms. When a autonomous machine is equipped with various means of social reinforcement in combination with algorithms that allow for adaptations to individuals, there is a greater chance of finding a reward that is of high subjective value for each participant. This can only be done if the machine is first equipped with these tools and then programmed in a way that allows it to perceive, adapt, and adjust rewards based on the feedback from participant progress. Mutual reinforcement is therefore a promising approach to skill transfer between a robot and human.
The following sections illustrate Algorithm 1 and the notations associated with it. It also explains the choice of selection of certain parameters and their impact in the experiment.
3.1. Optimization of Reinforcers
The above algorithm is implemented in both robot and computer gaming platforms inducing MRL. This method is directly applied to the problem of searching for an appropriate reward channel preferred by individual human participants during the skill transfer task. Here each reinforcer is influenced by the participant’s cognition and performance, and this evaluation directs which reinforcer will be considered next. Hence, in this method, the robot and the human are successively generating and evaluating attempts to obtain incremental improvements for each other. Fig 1 refers to the MRL concept where the participant rectifies the mistake after getting a positive reinforcer.
In the particular studies reported here, Baxter has four distinct reinforcers at its disposal, whereas for Tetris there are seven.
denotes the weight vector assigned to
reinforcers, which are initially a uniform probability distribution summing to 1. The reinforcers are given out on the basis of weighted random selection to meet the exploration-exploitation criteria, but since all the reinforcers are uniformly weighted at the beginning, weighted reinforcement in the first step is of no significance. If the reinforcergiven out from the set is a success then . Here represents the particular reinforcer that is provided by the robot or the machine to motivate people. is a small positive fraction called the step-size parameter, which influences the rate of learning. The value of is chosen empirically based on the observed performance of exploration and exploitation. Figure 2 denotes the entropy of the system with the suitable alpha value. Selection criteria for the value of are discussed in detail in the next subsection. The value of adds to the present weight of the reinforcer denoting its success. denotes the number of remaining reinforcement strategies, and the value of is equally distributed among them. If is successful, is added to it and is subtracted from all the rest of the reinforcers. In contrast, if is a failure, then is subtracted from it and is added to the rest of the reinforcers.
After this step the weights of
are updated. The mutual information shared among them is obtained from several interactions and the values of the reinforcers get updated every time with probabilities associated with faster skill transfer. The robot tries to learn about the person’s behavior and performance level and then applies this knowledge to motivate the individual. We used an exponential weighted moving average (EWMA) to gain information about the most recent interactions. For Baxter, since the number of reinforcers is fewer, we used the past three interactions and for Tetris we used five. EWMA only provides information about recent interactions, but we need to understand the variability of the reinforcers’ success over a longer term. Hence we maintain the value of two standard deviationsover the EMWA values in order to better notice and interpret success. Then again the probabilities of all the reinforcers in the set are updated and prepared for the next interaction. The robot stops giving out the reinforcers when the participant stops making mistakes.
MRL, an implementation of which we have described in the above algorithm, is a novel concept in the field of traditional reinforcement learning and can be implemented in several algorithmic approaches to get significant results.
3.2. Choice of Parameters
In Algorithm 1 we used several parameters whose values are tailored depending on the experimental requirements. The parameter is chosen on an empirical basis. We used numbers ranging from 0.01 to 0.07 in case of Tetris and 0.01 to 0.04 in case of Baxter. The range of numbers are selected on the basis of the number of reinforcers used in the experiment. We calculated entropy depending on the weight of to determine the mutual information gain over interactions and whether the autonomous agent is optimally exploring and exploiting. Using the maximum entropy principle, we know (Shannon, 1948) that entropy reduces over time with the information gain. We calculated the entropy using values chosen from the ranges given above to determine the most suitable value.
In Figure 2 we show various rates of entropy decrease for different values. The lines in lighter color show different values we considered for Baxter and Tetris and and the dark blue line exhibits the one we selected for our experiments, namely 0.015 for Baxter and 0.05 for Tetris. These values not only gain information linearly over time but also trade off satisfactorily between exploration and exploitation. In the process of information gain, the rate of decrease of entropy is not always perfectly linear; all the graphs of different values are accompanied by spikes due to exploration-exploitation tradeoffs. We designed the above algorithm not to behave greedily because exploring different reinforcers is necessary while breaking the monotony of the task. From these figures we can see that at the beginning of the experiment the entropy is maximum for all values and gradually decreases over interactions.
The value of is selected in such a fashion that the robot considers the last three interactions of Baxter and five for Tetris. is denoted as the multiplier. In our previous work (Roy et al., 2018a), we used the robot’s experience from the beginning to the task for the reinforcer selection and found out that sometimes people prefer more than one reinforcer. We theorized that their preference might have changed over interactions, and hence focused on recent interactions for better performance.
3.3. MRL and cognitive models
In traditional reinforcement learning designing an appropriate reward signal is a critical part of the application process. Various researchers have coined novel techniques to solve this issue (Abbeel and Ng, 2004). In contrast, processes like inverse reinforcement learning (IRL) learn from a expert’s behavior, where an agent tries to infer the reward signal to achieve a particular goal. In neither of these cases does any two-way interaction between the expert and the agent take place, and therefore they do not gain the advantage of situational feedback which is important during a learning process. To achieve a particular task, both expert and novice both should exchange feedback through appropriate reward channels. In practice, designing appropriate reward signals is often an informal trial-and-error search for a reward signal that produces acceptable results. In MRL, the expert explores and exploits the reward signals in the course of judging the novice’s actions and performance while accomplishing the task. Hence if the novice learns slowly, fails to lean or learns incorrectly the expert cooperates to improve the student’s learning during the process. This is a sophisticated way to find good reward signals, since feedback is given while accomplishing a subgoal and the expert can slowly guide the agent towards the overall goal. Hence unlike other reinforcement learning strategies, MRL is a complete model that supports task learning with human-robot interaction simultaneously learning about the reward preferences. In MRL, since the expert cooperates with the novice during the learning process, it also becomes aware of the cognitive models involved, which in turn leads to the design of better reward signals. To explore the efficiency of the process, we calculated the machine’s regret and the mutual information shared between the agents using Shannon’s entropy . Regret is defined as the difference between the reinforcer with maximum weight and the reinforcement strategy selected, i.e. .
Property 1: In MRL, an autonomous agent fails to identify the cognitive orientation of a participant if it crosses time steps as after steps no change in entropy occurs, which means no information gain.
The above matrix is a transition matrix with state space , where is possibly infinite. denotes the number of reinforcers used and and are the weights that are added to the system depending on the exploration-exploitation tradeoff. The above matrix is prepared on the basis of Algorithm 1. Now let be a row vector denoting a probability distribution on : so each element denotes the probability of being in state , and , where for all . The probability distribution is an equilibrium distribution for the above matrix if . That is, is an equilibrium distribution if
for all .
That is, , will have the same values and so on. This is because the values achieve numeric stability after . That means converges to a fixed matrix with all rows equal as . At this point, no further change in the Shannon entropy for , will be observerd. Entropy in a system denotes its information gain: a decrease in entropy means more information gain. Here has the maximum entropy 1 which decreases linearly over time with information gain. Now if does not change with time , that means the robot is not gaining any further information about the participant’s cognitive orientation. Hence we can conclude that cognitive orientation of a participant can only be found steps in the transition matrix.
Property 2: If MRL converges to stationary distribution over time (equilibrium), it is independent of the reinforcers used during the interaction.
From property 1, we determined the convergence criteria of . Hence we know that matrix converges to , for a large value of . Now the stationary or equilibrium distribution can be found out by solving
, where I is the identity matrix. If a matrix M reaches equilibrium at, we know that the cognitive orientation of the candidate is undetermined. Hence we assume that when equilibrium is achieved all the reinforcers are utilized and they failed to affect the human cognition. Hence we can conclude that if MRL converges to the equilibrium distribution, then it is independent of the reinforcers used during the interactions.
The autonomous agents weigh several positive reinforcers in this research to motivate the students if they commit any mistakes (Roy et al., 2019b). In this section we discuss the reinforcers used by the robot and the computer during the experiment and the effectiveness of Simpson’s psychometric model.
4.1. Reinforcers used with Baxter
When mistakes are made, Baxter forms a sad face and gives out the positive reinforcer to encourage participants (Fitter and Kuchenbecker, 2016), and when they perform correctly after the correction it forms a smiling face. Other than that, Baxter maintains a neutral face throughout the task. The following reinforcers are given out depending upon the subject’s assigned experimental group.
Verbal reinforcer : When using this reinforcer, the robot asserts that it is trying to encourage the subject with some positive feedback. Since Baxter does not have its own audio interface, we used speakers to produce the robot voice. In our experiment, if the subject makes a mistake, the robot will verbalize something like, “Sorry dear, don’t worry. You can do it”.
Hint-based reinforcer : This takes the form of a hint given to the participant during a task. The hint does not provide the correct answer but tries to influence the subject’s thought process so that it increases the learning rate of the participant. For example, during the pattern making process, if a candidate places an incorrect marker, Baxter suggests flipping the marker box and trying the other side, before rejecting the block entirely. Thus people can track the blocks they have already tried to place in a particular spot.
Simple-feedback reinforcer: In this case, the robot only identifies the correct or the incorrect marker. It doesn’t attempt to induce any kind of positivity or motivation in the participant. This is because some people are not fond of external motivations and only a rectification in the task can influence them to perform better.
Gesture-based reinforcer: In this case, the robot adds a consoling gesture by patting at the student’s back and also provides positive verbal feedback as referenced above.
4.2. Reinforcers used with Tetris
In the case of Tetris, seven different positive reinforcers are used during the interactions. We increased the number of reinforcers because in a fast-moving gaming scenario, we anticipated more interactions per session. All the positive reinforcers are displayed in an audio-visual setting whenever a participant makes a mistake. Here all of the reinforcers provided some sort of hint for the player to perform better. For example, whenever a player is playing for too long without scoring any points, reinforcers are provided such as “Clear the lines quickly for faster score” or guiding the player to check for the upcoming blocks to plan the next move ahead. The type of incentives are manipulated according to the platforms we employed to demonstrate the effectiveness of MRL.
4.3. Simpson’s psychometric model
is used to identify the skill transfer. This model characterizes the principles of skill evaluation, which are closely linked with important aspects of human cognition. It is widely used by teachers, professional specialists and scientists to evaluate curricular problems with greater precision. Simpson’s psychometry domain is defined in form of a taxonomy which gives us a clear idea about how knowledge is acquired by an individual and how that is later applied to execute tasks. Simpson’s psychometric model is broadly classified into Perception, Set, Guided Response, Mechanism, Complex Overt Response, Adaptation and Origination. Perception is related to the awareness of the present situation. Set is the eagerness of the human participant to volunteer for the task. Guided Response is the early stage of learning a complex skill with the help of an instructor. After the participant has learned the task, the later stages of the psychometric domain involve applying the training. Mechanism is the immediate step to demonstrate basic proficiency with respect to a simple application. Complex Overt Response is associated with skillfully applying complex versions of the same task with greater proficiency. Adaptation signifies complete learning, where individuals can respond to uncertain events, while Origination is the last phase of learning where humans can generate new ideas from their knowledge. Here the experiments are designed to confirm the feasibility of Simpson-based skill evaluation. Since the tasks are designed in a lab setting, we only used a few of the above categories to determine the skill transfer process. The following section discusses the experimental models and the findings associated with them.
5. Task description with Baxter
(age : , , male=13, female=21, none=11, random=11, learned=12) participants were recruited for the experiments with Baxter, which ran for a time (, in minutes). Among the subjects, 75% had never interacted with a robot, 8.30% interacted a year ago, and 5.50% each a month and a week ago. The task involved in the experiment was divided into two large sections that was further divided into two smaller subsections. The tasks are designed to observe successful skill transfer from robot to human using Simpson’s psychometric model, where each category in the taxonomy transitions to another with the goal of skill transfer. In this experiment we only used Guided Response, Mechanism and Adaptation. Guided Response I and II occur in the first half of the experiment where the robot first teaches the participant about the augmented markers and then motivates them throughout the learning process. During this the robot also evaluates the performance of the participants. In the second half of the first section the robot teaches the participant a complex pattern with dual-faced augmented markers and asks the student to reconstruct it. Again, during this process, the robot positively reinforces the learner with simple yes-or-no, random or learned MRL feedback depending upon their assigned experimental group. Participants are allowed to observe the pattern making process and then the markers are immediately shuffled and they are asked to start the reconstruction immediately. Baxter transitions its left hand camera from one spot to another for evaluation. Baxter does not progress to the next position until the participant rectifies any mistakes. During this process the participants get hint, simple, verbal or gesture feedback depending upon their group and Baxter with their performance tries to identify their cognitive orientation. Fig 3 (top frames) corresponds to the experimental procedure with Baxter.
In the second half of the experiment the participants are asked twice to reconstruct the pattern, this time without any motivation, to observe how well the skill transfer succeeded. In the last half of the experiment participants are asked again to identify two random markers from the set they were taught at the beginning of the experiment to analyze adaptation. Each participant in the experiment was assigned different complex patterns for reconstruction. The marker placement at each position depends upon the robot. At the end of the experiment, the subjects were given questionnaires to answer using a 5-point Likert scale.
The results section is further divided into subsections discussing the performance of the participants, the mutual information shared, the robot’s regret and the mental model of the participants during the task.
5.2. Subjective performance evaluation
To quantify the skill transfer procedure, we calculated the number of mistakes made by each participant in all of the groups in different phases of the experiment. Figure 4 shows the number of mistakes according to Simpson’s psychometric model. There are participants () who didn’t commit any mistake throughout the experiment, so their results are not included in the mistakes data. From Table 1 and Figure 4 we can see that the number of mistakes made during the Mechanism and Adaptation phases are significantly less than during the Guided Response phases irrespective of the groups, which shows the effect of robot feedback during the task. Again if we compare the skill transfer among the groups in the figure, we can see that the number of mistakes in the learned MRL group is comparatively lower than the other two. Table 1 presents the performance of the participants in all the phases of the experiment. Figure 4
also plots the number of mistakes in each group before (Guided Response I and II) and after (Mechanism and Adaptation) the skill transfer procedure; we see that the participants made comparatively fewer mistakes in the learned group than in the other two. Using a t-test for skill transfer outcomes while comparing the learned group with random reinforcers we get a significantvalue 0.05. Thus we can conclude that MRL has successfully worked in terms of the skill transfer procedure.
|Guided Response I||0.4||0.84||0.8||1.17||2.08||3.92|
|Guided Response II||2.08||3.92||7.45||8.37||4.5||4.80|
5.3. Entropy analysis
Entropy denotes the randomness of a system. In Figure 5, we can see that the entropy of the information of the robot obtained by interactions goes down monotonically along with each interaction with its human participant in MRL. With each interaction, the robot is gaining more information about the participants’ performance and their cognitive orientation towards each reinforcer. The pale green lines show the entropy of the information of the each participant obtained by interactions with Baxter, and the blue line is their mean performance. Not every participant had an equal number of interactions with Baxter, but regardless, using Algorithm 1, the machine manages to gain information steadily about their performance. For each participant, the value of entropy varied depending on the exact pattern of the robot’s choices of exploration and exploitation, but we can conclude that for each participant it has gained some information at each interaction which later helped it to construct a successful mental model.
5.4. Regret analysis
As mentioned in Section 3.3, regret is calibrated on the basis of the decision making ability of the robot. It depends upon the subject’s performance, which helps in characterizing the most appropriate reinforcement learning strategy. We correlated the number of mistakes made by the participants and the total regret felt by the robot (Figure 6) and found a linear relationship between the number of mistakes and the robot’s regret. The value of the coefficient is ; thus the robot’s regret is strongly correlated and the reinforcement learning strategy used by the robot to understand human responses and improve their performance, is working appropriately. For the participants who had several interactions with the robot or made many mistakes, Baxter tried to explore different reinforcement strategies at different times, trying to increase their learning rate. This illustrates that Baxter can successfully train people to achieve complex task using their preferred motivations.
5.5. Mental model analysis
At the end of the experiment, those participants in the learned model group were asked to choose their preferred reinforcers. Baxter could correctly identify the preferred reinforcers in half of the cases (twice as effectively as a random baseline). Thus, MRL allowed the robot successfully to identify the cognitive orientation of the participants to a large extent (accuracy score 0.50). During the task, since the number of interactions was limited, the robot did not have sufficient opportunity to engage in the exploitation aspect of the reinforcement learning, and thus its ability to identify preferred reinforcers was limited (but still reasonably successful). In these experiments, Baxter explored more than exploited, which impacted the types of reinforcers given out by the robot.
At the end of each experiment we probed participants with a 5-point Likert scale (1: Strongly disagree; 2: Disagree; 3: Neutral; 4: Agree; 5: Strongly agree). Participants in the no reinforcer group (=3.54 , =0.68 ), random reinforcer group (=4 , =0.77 ), learned reinforcer group (=3.25 , =1.22 ) wanted to play with Baxter again and again in no reinforcer group (=4.0 , =0.63), random reinforcer group (=3.63 , =0.92) and learned reinforcer group (=2.6 , =0.89) thought it is useful as a teacher. Interestingly, the people who performed poorly during the experiment neither found Baxter to be useful nor thought it was a good teacher. In other words, people who enjoyed the interaction also found it helpful and wanted to come again to learn from the robot, whereas those were not fond of the robot were not interested in the experiment and ended up performing poorly.
6. Task Description with Tetris
(age , , male=9, female=22, none=11, random=10, learned=10) participants were recruited for the experiments with Tetris for 15 minutes. We conducted an experiment in a gaming scenario to observe the performance of MRL across different platforms. Among all the participants, of the subjects had played the game but not within the last year, had played within the last year but not the last month, and the rest had played the game within the last month. Like with the Baxter scenario, participants were also being trained to be better Tetris players. The skill transfer scenario is also analyzed with Simpson’s psychometric model (Guided Response and Adaptation). As is common in Tetris games, during each move, the next block is shown alongside the 10x20 game board so that players plan their move ahead. As in the previous experiment, the teaching process is divided into two phases. Initially the participants are asked to play the game for 15 mins with reinforcers provided depending upon their assigned experimental group (simple, random, MRL).
Whenever the participant makes a mistake, the machine alerts them, provides a reinforcer, and then they are allowed to continue. For Tetris, mistakes are considered to be placing a block in such a way to hinder fast scoring. Wrong placement is associated with forming a gap between lines which will make it difficult to eliminate the line of blocks in the future. Also, if a player places blocks several times in a row without eliminating any lines, that is also considered as a mistake as the towers of blocks build up and get closer to ending the game. During the first experiment task, if the game is lost, participants are allowed to restart, so that all subjects received the same time to learn properly. In the second portion of the experiment (Adaptation), participants are asked again to play the game, this time without reinforcement training, to assess the quality of their learning. They are asked to play until losing, or until 5 minutes elapsed (=2.68 =0.88 minutes). Since the experiment had only two phases, we used Guided Response and Adaptation to analyze the skill transfer. After the game, the subjects were also given questionnaires probed with a 5-point Likert scale.
Similarly to the Baxter experiment, our analysis of the Tetris scenario is divided into subjective and entropy evaluation, regret and mental model analysis.
6.2. Subjective performance evaluation
For Tetris, we computed the skill transfer on the basis of Simpson’s psychometric model. Since, in a game like Tetris, people are expected to make a large number of mistakes, we computed the scores per minute of the participants of different groups. Figure 7 denotes the scores of the participants before (Guided Response) and after (Adaptation) the skill transfer procedure. The subjects in the learned group scored significantly better than the other two (). Although reinforcers were provided in the random group, they were not tailored to individual learning styles, and learners responded better to MRL. Subjects performed conclusively better with MRL than other two groups.
6.3. Entropy analysis
As shown in Figure 8, the entropy of the gaming device’s information monotonically decreases with each interaction with its human participant in MRL. Like Baxter, this shows that the computer is gaining information about the participant’s performance and its cognitive orientation towards each reinforcer with each interaction. The pale green lines show the entropy of each participant over each interaction with the machine and the blue line shows the mean performance.
6.4. Regret analysis
Like Baxter, here also we tried to calibrate the regret of the gaming system. The correlation coefficient in this case is . Although the system tried to explore different reinforcement strategies at different times, trying to increase the learning rate, the results are suggestive and not conclusive. Fig. 9 is the best fit curve to analyze the relation between regret of the system and the integral of the number of mistakes made by the participants. This is because the participants utilized all the reinforcers given to them without distinguishing between them very much.
6.5. Mental model analysis
At the end of the experiment, those participants in the learned model were asked to choose their preferred reinforcers. The machine could partially identify the preferred reinforcers in Tetris. For example, reinforcer a is often misclassified as e in many cases. In Tetris, since all the reinforcers were fairly similar to one another, in spite of the fact that they guided participants differently, the subjects often failed to distinguish the efficacy of one versus another(accuracy score 0.50). Hence they were able to make use of all of the provided reinforcers successfully, without much differentiation between them. It was difficult for the system to determine the preferred reinforcer of the participant, and only it occasionally successfully identified the best reinforcers.
At the end of each experiment we probed participants with a 5-point Likert scale (1: Not all helpful; 2: Moderately helpful; 3: Neutral; 4: Helpful; 5: Extremely helpful). The subjects wanted to play Tetris again with (=3.0, = 0.70) none, (=3.2, = 0.78) with random and (=3.3, = 0.82) with MRL. They thought it was useful (=2.1, = 0.75) with none, (=2.5, = 0.85) with random and (= 2.5, =0.97) with MRL. Overall, more than half the subjects found the reinforcers useful and therefore thought the system was a good teacher. These statistics show significant results.
In the experiment, the results are sometimes not as strong as we might hope for several reasons. If the robot’s grippers were closed, they occasionally hindered the camera, blocking the robot from identifying the markers, since we used the left hand camera for detection and evaluation purposes. The Baxter arm and gripper are not extremely dextrous; it is sometimes very unsophisticated in its attempts to pick up the blocks whenever they not lying perpendicularly to the camera. The markers, after several tasks and the degradation which resulted from repeated handling by both human and robot, became unclear and difficult for the marker tracking algorithm to recognize, which also contributed to system crashes. Again, many of the young adults who participated in the experiment failed to connect to the robot emotionally and lacked engagement. Some subjects paid very little attention to the robot’s attempts to communicate a reinforcement strategy, to the point that the subjects attempted to interact with the researchers conducting the experiment rather than the robot. Some participants simply produced iteration after iteration of patterns until they happened upon the correct one, without paying attention to the robot gamely attempting to help. Rather, they simply tried each block at each position to figure out the right approach. Hence Baxter on its end was confused in providing the reinforcement strategy. For this reason, we see that Baxter was only able to identify a successful motivational strategy for half the participants in the learned group.
Another potential point of alienation came from the fact that Baxter’s voice did not issue from the robot itself, but rather a speaker off to the side (since the robot hardware lacks sound capability). The students had to turn to their right side and interact with a computer console to give Baxter their feedback in form of yes or no, which is unsophisticated; verbal interaction would have been a better option. However, participants in the experiment came from different national backgrounds and language abilities, so it was very hard for the robot to understand their pronunciation, and we were therefore forced to keep the human feedback in that format.
Seven blocks with 14 markers can be placed in many, many ways, but no participants required nearly that many attempts to figure out the correct pattern. Thus, even when they did not directly engage with the robot’s attempts to teach, it still had some impact on them. We used the spatial arrangement of markers as our complex task. Some people who performed better in all the groups might simply be good at this style of task and would fail at some other complex task. The difference in performance between the groups might be different in different complex task scenarios. Our MRL theory applies to the people who have performed poorly in the task, and therefore received appropriate motivations. They might be better at a different complex task, and therefore engage differently with the reinforcement behavior. In the case of Tetris, people are well acquainted with the game, so the reinforcers might only have a little effect on them. Although the autonomous agents successfully developed a teaching strategy for only half of the participants, it is enough to suggest that such feedback does have impact on human behavior and learning. Furthermore, this approach allows the autonomous agent to assess its own success and learn to calibrate its own interactions in ways that lead to successful teaching.
7. Conclusion and future work
In this work, we studied the problem of skill transfer from a robot to human, where the autonomous agent is not only learning about the human mental model but also trying to adapt its own accordingly. In this shared environment, the robot is trying to maximize the cumulative reward by learning about human behavior and simultaneously improving its own cognitive model. We highlighted mutual information communicated between the robot and the human, and validated their interaction in skill transfer using real-time experiments in both robot and gaming platforms. The subjective performance, information gain over time and the confusion matrices give us a conclusive idea how robots and computer systems can successfully transfer skills from themselves to humans.
In our future work we would like to implement MRL across different platforms. Heavy construction equipment like excavators and backhoes are required to perform complex tasks like digging, truck loading and ditch crossing, requiring a series of complex manipulations. Learning appropriate manipulations for these different situations is a hard task. We want to implement MRL in these scenarios where humans can learn the subtlety of control manipulations with robot assistance. In addition, we intend to investigate the necessary behavioral changes required to be adapted by the robots to become better trainers over time. We would also like to involve robots in guiding students towards correct actions and the responses they should develop, keeping the learners more thoroughly engaged in the task. We want to design and identify suitable reward channels for individuals to learn a task proficiently, simultaneously identifying their mental models in a robot-human interaction framework.
Acknowledgements.This work was supported by NSF award #1527828 (NRI: Collaborative Goal and Policy Learning from Human Operators of Construction Co-Robots).
Abbeel and Ng (2004)
Pieter Abbeel and
Andrew Y. Ng. 2004.
Apprenticeship learning via inverse reinforcement
Proceedings of the 21st International Conference on Machine Learning (ICML).
et al. (2009)
Rémi Munos, and Csaba
Exploration–exploitation tradeoff using variance estimates in multi-armed bandits.Theoretical Computer Science 410, 19 (2009), 1876–1902.
- Bickhard (2009) Mark H Bickhard. 2009. Interactivism: a manifesto. New Ideas in Psychology 27, 1 (2009), 85–95.
- Boccanfuso et al. (2016) Laura Boccanfuso, Erin Barney, Claire Foster, Yeojin Amy Ahn, Katarzyna Chawarska, Brian Scassellati, and Frederick Shic. 2016. Emotional robot to examine different play patterns and affective responses of children with and without ASD. In 2016 11th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 19–26.
- Bradley Knox and Stone (2008) W. Bradley Knox and P. Stone. 2008. TAMER: Training an Agent Manually via Evaluative Reinforcement. In 2008 7th IEEE International Conference on Development and Learning. 292–297. https://doi.org/10.1109/DEVLRN.2008.4640845
- Cakmak et al. (2009) Maya Cakmak, Nick DePalma, Andrea L Thomaz, and Rosa Arriaga. 2009. Effects of social exploration mechanisms on robot learning. In RO-MAN 2009-The 18th IEEE International Symposium on Robot and Human Interactive Communication. IEEE, 128–134.
- Chan et al. (2019) Lawrence Chan, Dylan Hadfield-Menell, Siddhartha Srinivasa, and Anca Dragan. 2019. The Assistive Multi-Armed Bandit. In 2019 14th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 354–363.
- Cheung and Gardner (2015) Elaine O Cheung and Wendi L Gardner. 2015. The way I make you feel: Social exclusion enhances the ability to manage others’ emotions. Journal of Experimental Social Psychology 60 (2015), 59–75.
- Dominey and Warneken (2011) P. F. Dominey and F. Warneken. 2011. The basis of shared intentions in human and robot cognition. New ideas in psychology. 29, 3 (2011).
- Fan et al. (2018) Yang Fan, Fei Tian, Tao Qin, Xiang-Yang Li, and Tie-Yan Liu. 2018. Learning to teach. arXiv preprint arXiv:1805.03643 (2018).
- Fasola and Matarić (2013) Juan Fasola and Maja J Matarić. 2013. A Socially Assistive Robot Exercise Coach for the Elderly. J. Hum.-Robot Interact. 2, 2 (June 2013), 3–32. https://doi.org/10.5898/JHRI.2.2.Fasola
- Ferreira and Lefèvre (2015) Emmanuel Ferreira and Fabrice Lefèvre. 2015. Reinforcement-learning Based Dialogue System for Human-robot Interactions with Socially-inspired Rewards. Comput. Speech Lang. 34, 1 (Nov. 2015), 256–274. https://doi.org/10.1016/j.csl.2015.03.007
- Fitter and Kuchenbecker (2016) Naomi T Fitter and Katherine J Kuchenbecker. 2016. Designing and assessing expressive open-source faces for the Baxter robot. In Proceedings of the International Conference on Social Robotics (ICSR).
- Griffith et al. (2013) Shane Griffith, Kaushik Subramanian, Jonathan Scholz, Charles L Isbell, and Andrea L Thomaz. 2013. Policy shaping: Integrating human feedback with reinforcement learning. In Advances in neural information processing systems. 2625–2633.
- Hadfield-Menell et al. (2016) Dylan Hadfield-Menell, Stuart J Russell, Pieter Abbeel, and Anca Dragan. 2016. Cooperative inverse reinforcement learning. In Advances in neural information processing systems. 3909–3917.
- Keskin (2014) Sevgi Coşkun Keskin. 2014. From what isn’t empathy to empathic learning process. Procedia-Social and Behavioral Sciences 116 (2014), 4932–4938.
- Kim et al. (2013) Elizabeth S Kim, Lauren D Berkovits, Emily P Bernier, Dan Leyzberg, Frederick Shic, Rhea Paul, and Brian Scassellati. 2013. Social robots as embedded reinforcers of social behavior in children with autism. Journal of autism and developmental disorders 43, 5 (2013), 1038–1049.
- Knox and Stone (2009) W. Bradley Knox and Peter Stone. 2009. Interactively Shaping Agents via Human Reinforcement: The TAMER Framework. In Proceedings of the Fifth International Conference on Knowledge Capture (K-CAP ’09). ACM, New York, NY, USA, 9–16. https://doi.org/10.1145/1597735.1597738
- Koenig et al. (2010) N. Koenig, L. Takayama, and M. Mataric. 2010. Communication and knowledge sharing in human–robot interaction and learning from demonstration. IEEE transactions on neural networks 23, 8-9 (2010).
- Konok (2018) Veronika Konok. 2018. Should we love robots?–The most liked qualities of companion dogs and how they can be implemented in social robots. Computers in Human Behaviour 80 (2018), 132–142.
- Lazarescu (2013) Mihaela Paisi. Lazarescu. 2013. The Structure and Dynamics of the Teacher’s Empathic Behavior. Procedia-Social and Behavioral Sciences 78 (2013), 511–515.
- Lee et al. (2005) Sau-lai Lee, Ivy Yee-man Lau, Sara Kiesler, and Chi-Yue Chiu. 2005. Human mental models of humanoid robots. In Proceedings of the 2005 IEEE international conference on robotics and automation. IEEE, 2767–2772.
- Leite et al. (2012a) Iolanda Leite, Ginevra Castellano, André Pereira, Carlos Martinho, and Ana Paiva. 2012a. Long-term interactions with empathic robots: Evaluating perceived support in children. In International Conference on Social Robotics. Springer, 298–307.
- Leite et al. (2012b) Iolanda Leite, André Pereira, Samuel Mascarenhas, Carlos Martinho, Rui Prada, and Ana Paiva. 2012b. The influence of empathy in human–robot relations. International Journal of Human-Computer Studies 71 (01 2012). https://doi.org/10.1016/j.ijhcs.2012.09.005
- Litoiu and Scassellati (2015) Alexandru Litoiu and Brian Scassellati. 2015. Robotic coaching of complex physical skills. In Proceedings of the Tenth Annual ACM/IEEE International Conference on Human-Robot Interaction Extended Abstracts. ACM, 211–212.
- Lockwood (2016) Patricia L Lockwood. 2016. The anatomy of empathy: Vicarious experience and disorders of social cognition. Behavioural brain research 311 (2016), 255–266.
- Lopes et al. (2009) Manuel Lopes, Francisco Melo, and Luis Montesano. 2009. Active learning for reward estimation in inverse reinforcement learning. In Proceedings of the Joint European Conference on Machine Learning (ECML).
- Lopes et al. (2013) Manuel Lopes, Didier Roy, Pierre-Yves Oudeyer, and Benjamin Clement. 2013. Multi-armed bandits for intelligent tutoring systems. arXiv preprint arXiv:1310.3174 (2013).
- Nikolaidis and Shah (2013) Stefanos Nikolaidis and Julie Shah. 2013. Human-robot cross-training: computational formulation, modeling and evaluation of a human team training strategy. In Proceedings of the 8th ACM/IEEE international conference on Human-robot interaction. IEEE Press, 33–40.
- Nunez et al. (2015) Eleuda Nunez, Soichiro Matsuda, Masakazu Hirokawa, and Kenji Suzuki. 2015. Humanoid robot assisted training for facial expressions recognition based on affective feedback. In International Conference on Social Robotics. Springer, 492–501.
- Park et al. (2017) Hae Won Park, Rinat Rosenberg-Kima, Maor Rosenberg, Goren Gordon, and Cynthia Breazeal. 2017. Growing growth mindset with a social robot peer. In 2017 12th ACM/IEEE International Conference on Human-Robot Interaction (HRI. IEEE, 137–145.
- Powell and Roberts (2016) Philip Powell and Jennifer Roberts. 2016. Situational determinants of cognitive, affective, and compassionate empathy in naturalistic digital interactions. Computers in Human Behaviour 68 (11 2016). https://doi.org/10.1016/j.chb.2016.11.024
- Prozesky (2000) Detlef R Prozesky. 2000. Teaching and learning. Community eye health 13, 36 (2000), 60.
- Ramachandran and Scassellati (2015) Aditi Ramachandran and Brian Scassellati. 2015. Developing adaptive social robot tutors for children. In 2015 AAAI Fall Symposium Series.
- Roy et al. (2018a) Sayanti Roy, Christopher Crick, Emily Kieson, and Charles Abramson. 2018a. A reinforcement learning model for robots as teachers. In 2018 27th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN). IEEE, 294–299.
- Roy et al. (2018b) Sayanti Roy, Emily Kieson, Charles Abramson, and Christopher Crick. 2018b. Using human reinforcement learning models to improve robots as teachers. In Companion of the 2018 ACM/IEEE International Conference on Human-Robot Interaction. ACM, 225–226.
- Roy et al. (2019a) Sayanti Roy, Emily Kieson, Charles Abramson, and Christopher Crick. 2019a. Effective robot-human skill transfer via mutual reinforcement learning. (2019).
- Roy et al. (2019b) Sayanti Roy, Emily Kieson, Charles Abramson, and Christopher Crick. 2019b. Mutual Reinforcement Learning with Robot Trainers. In 2019 14th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 572–573.
- Salter et al. (2006) Tamie Salter, Kerstin Dautenhahn, and Rene Boekhorst. 2006. Learning about natural human-robot interaction styles. Robotics and Autonomous Systems 54(2) (02 2006), 127–134. https://doi.org/10.1016/j.robot.2005.09.022
- Sauppé and Mutlu (2015) Allison Sauppé and Bilge Mutlu. 2015. Effective Task Training Strategies for Human and Robot Instructors. Auton. Robots 39, 3 (Oct. 2015), 313–329. https://doi.org/10.1007/s10514-015-9461-0
- Scassellati (2001) Brian M Scassellati. 2001. Foundations for a Theory of Mind for a Humanoid Robot. Ph.D. Dissertation. Massachusetts Institute of Technology.
- Shannon (1948) Claude Elwood Shannon. 1948. A mathematical theory of communication. Bell system technical journal 27, 3 (1948), 379–423.
- Simpson (1972) E Simpson. 1972. The classification objectives in the psychomotor domain–Washington. Gryphon House (1972).
- Simpson (1966) Elizabeth J Simpson. 1966. The Classification of Educational Objectives, Psychomotor Domain. (1966).
- Spaulding et al. (2018) Samuel Spaulding, Huili Chen, Safinah Ali, Michael Kulinski, and Cynthia Breazeal. 2018. A Social Robot System for Modeling Children’s Word Pronunciation: Socially Interactive Agents Track. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 1658–1666.
et al. (2006)
Georgi Stojanov, Goran
Trajkovski, and Andrea Kulakov.
Interactivism in artificial intelligence (AI) and intelligent robotics.New ideas in psychology. 24, 2 (2006).
- Sutton and Barto (2018) Richard S Sutton and Andrew G Barto. 2018. Reinforcement learning: An introduction. MIT press.
- Tabrez et al. (2019) Aaquib Tabrez, Shivendra Agrawal, and Bradley Hayes. 2019. Explanation-based reward coaching to improve human performance via reinforcement learning. In 2019 14th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 249–257.
Yi et al. (2018)
Xiaoyuan Yi, Maosong Sun,
Ruoyu Li, and Wenhao Li.
Automatic Poetry Generation with Mutual
Reinforcement Learning. In
Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 3143–3153.
- Yin et al. (2015) Hang Yin, Aude Billard, and Ana Paiva. 2015. Bidirectional learning of handwriting skill in human-robot interaction. In Proceedings of the Tenth Annual ACM/IEEE International Conference on Human-Robot Interaction Extended Abstracts. ACM, 243–244.