Model Player or Model-Free refers to, in the context of the provided paper, to whether or not the underlying Markov Decision Process is learned during training or not. It has nothing to do with classifiers/regressors in general or meta-heuristics or anything like that. The authors are discussing Reinforcement Learning. A regressor can be used to Learn Q-Values (as in Q-Learning, SARSA - what you call "cached" values) or the Model itself. If you do Model-Free learning, your agent learns to act without reference to a model (i.e. the underlying world dynamics). If you do model-based learning, you first learn the model and base your action on that model. Methods that learn using the model are Value Iteration/Policy iteration, R_max (Brafman & Tennenholtz 02)

See David Silver's slides here: http://www0.cs.ucl.ac.uk/staff/D.Silver/web/Teaching_files

And my Slides here:

http://nbviewer.ipython.org/github/ssamot/RL_notebook/blob/master/RL/RL.ipynb

AI techniques are mostly model free in the technical sense, but of course an AI design will inevitably be based upon certain prior principles and presumptions. AI is best seen, in my view, as an engineering discipline

What assumptions are AI technique needed? such as neural network, and support vector machine .

Some hyper-heuristics frameworks are not model-free. The very dependent to the learning or the problem to solve.

There is a new line of research that is focuses on improving this issue thought.

Conference Paper A Review of Hyper-Heuristics Framework

Conference Paper Plug-and-Play hyper-heuristics: an extended formulation

Model Player or Model-Free refers to, in the context of the provided paper, to whether or not the underlying Markov Decision Process is learned during training or not. It has nothing to do with classifiers/regressors in general or meta-heuristics or anything like that. The authors are discussing Reinforcement Learning. A regressor can be used to Learn Q-Values (as in Q-Learning, SARSA - what you call "cached" values) or the Model itself. If you do Model-Free learning, your agent learns to act without reference to a model (i.e. the underlying world dynamics). If you do model-based learning, you first learn the model and base your action on that model. Methods that learn using the model are Value Iteration/Policy iteration, R_max (Brafman & Tennenholtz 02)

See David Silver's slides here: http://www0.cs.ucl.ac.uk/staff/D.Silver/web/Teaching_files

And my Slides here:

http://nbviewer.ipython.org/github/ssamot/RL_notebook/blob/master/RL/RL.ipynb

There are two main techniques in AI, statistical base, syntactical base.

For statistical base techniques are generally free, but the results were less able to be controlled.

For syntactical base techniques is not free, depending on the class of problems that would be resolved, but the results

can be controlled in accordance with the riches of rules that can be provided.

In the context of reinforcement learning, there are some approaches that are model based, i.e., approaches that learn a model of the agent/environment. A key issue with model-based RL is that, if you don't take care, you suffer from model errors: If you believe that your learned model is 100% correct then you will make decisions based on this (unrealistic) assumption. And then RL typically runs into problems. 

My impression is that people tend to agree that you need to take care of this problem, e.g.,  by using probabilistic models, such as Bayesian linear models or Gaussian processes - at least in the case where you don't have enough data:

Schneider (NIPS 1997):

Exploiting Model Uncertainty Estimates for Safe Dynamic Control Learning

Deisenroth et al. (ICML 2011):

PILCO: A Model-Based and Data-Efficient Approach to Policy Search

Especially when you work with real systems (robots or so) you probably want to use model-based RL methods if somehow possible: They often require significantly fewer experiments with the robot than model-based approaches, which is important: experimentation time and hardware wear and tear should not be under-estimated.

Summarizing this comment:

- Model-based RL is not too widespread because learning models is hard (-> model errors are a big issue). Therefore, you should consider probabilistic models.

- Model-based RL is important in situations where you can't run thousands of experiments (e.g., robotics) to cut down experimentation time

http://papers.nips.cc/paper/1317-exploiting-model-uncertainty-estimates-for-safe-dynamic-control-learning.pdf

Conference Paper PILCO: A Model-Based and Data-Efficient Approach to Policy Search.

On Model-Free

Every computer program is a model (from a programmer’s mind). So what we might mean by model-free is that the program does not include a model. Model-free implies simple retrieval programs in which return “knowledge” (fixed output) from a given input. If the system modifies that output it is likely to do so from further input sources (including context). In which case it may be modelling.

On AI

The word intelligence implies adaptation and in turn implies modification essentially by feedback (learning). This suggests that any system in which feedback from both an external or internal source is a requisite for a property we call intelligence. If the response is internal then we can describe it as modelling. Thus for AI system which exhibit obvious intelligence there will be modelling. For a system externally adapting to feedback we might say they are model free if the system is automatically adjusting the paths to address appropriate responses (ie adaptive control). But such a system can be viewed as a pair of matched components (system + environment) and therefore can be described as model-pair (like a co-routine) - there is still the property intelligence but is difficult to see where it resides.

Thus, I contend that, any system which can adapt (possesses intelligence) is not model-free. To answer the question:

All AI systems are model-based there are no model-free AI systems.

Perhaps a better description would be : all AI system are control-based (loop: (sensor-input , feedback & modification, final response ) matched to a consistent environment/internal model of environment)?

I believe it model free techniques are hindrance than boon. First of all it is not possible to create a complete 100% model free system. An application is often valid for a certain class of problems, disciplines or subjects for which the knowledge is mined. 2nd the amount of effort time, energy and AI techniques such as neuron network and training required is often overwhelming in order to perfect a system. 

Even after doing so, the resulting system is very difficult to modify and maintain.

To take an historical perspective  - concepts like mapping, planning and representation were central to AI from the beginning - ie - Newell and Simon's Physical Symbol System Hypothesis to the Sense Map Plan Act paradigm of the 80s. AI 'gods' Minsky and Papert denied the 'intelligence' of neural network/connectionist approaches, arguing for 'symbols' and logical operations on symbols.The connectionists argued that neural nets learned and did not have internal representation - ie - models.

In AI, the major books included Schank's Scripts, Plans, Goals, and Understanding, or the work of David Marr in vision and visual representation. A symbol is a representation, and is thus is a model. This kind of work was rightly referred to by John Haugeland as GOFAI - Good Old Fashioned Artificial Intelligence. Then the famous Common Sense Problem occurred - formented by John Searle, who showed rhetorically - that models lead to brittleness - and he rise of situated and bottom up approaches,famously Rodney Brooks subsumption architecture, and generally by artificial life, situated and embodied approaches.  An anti representational movement began, - ie- lucy suchman in her Plans and situated actions (87)

Fastforward 25 years and we see that connectionist/neural networks have become part of AI. Is this a defeat for conventional AI or progress? Both I suppose.

So - your question has deep historical and philosophical roots. At issue is the basic question - is intelligence/thinking - in human, animal or machine -  fundamentally or necessarily representational. That is, is it fundamentally or necessarily internalist?